[ { "gpu_id": "a10-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "a10-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 35.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 34.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 29.8 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 25.6 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.5 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 35.3 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 29.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.0 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 30.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "40 GB of memory is below the 41.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.7 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 25.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "40 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -22.9 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.7 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.6 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "40 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "40 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.2 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-40gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "40 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 40, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.8 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 75.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 74.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 72.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 69.8 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 65.6 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 52.5 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 75.3 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 69.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 60.0 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 70.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "llama-3.1-70b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 50.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 29.3 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 65.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "gpt-oss-120b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 62.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.1 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 58.7 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 60.6 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 64.4 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 58.2 }, "rule_version": 1 }, { "gpu_id": "a100-pcie-80gb", "llm_id": "glm-4.5-air", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 74.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.2 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "a16-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "a30-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "a40-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "arc-a310-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "arc-a380-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "arc-a580-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "arc-a750-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "arc-a770-16gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "arc-a770-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.7 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.4 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.5 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.3 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "gemma-3-12b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 9.1 GB required at a context length of 4,096 tokens, but not the 10.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.0 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -40.7 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -52.9 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.3 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.8 }, "rule_version": 1 }, { "gpu_id": "arc-b570-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -64.8 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "arc-b580-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-1010-2gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-1030-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "gt-610-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "gt-620-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-630-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-640-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-710-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-730-2gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gt-730-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1050-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1060-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1070-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "11 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -39.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -51.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "11 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1080-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -63.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1630-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1650-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-6gb", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-6gb", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-1660-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "gtx-650-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-660-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-670-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-680-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-750-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-760-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-770-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.5 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "3 GB of memory is below the 3.9 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.2 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -24.5 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "3 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.7 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.0 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -47.7 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -59.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.3 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.4 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.8 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -71.8 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.5 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "3 GB of memory is below the 3.9 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.2 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -24.5 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "3 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.7 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.0 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -47.7 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -59.9 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.3 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.4 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.8 }, "rule_version": 1 }, { "gpu_id": "gtx-780-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -71.8 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-950-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "gtx-960-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-970-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-980-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-black-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "gtx-titan-x-12gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 75.1 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 74.4 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 72.1 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 69.8 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 65.6 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 52.5 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 75.3 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 69.1 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 60.0 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 70.4 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "llama-3.1-70b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 50.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 29.3 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 65.4 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "gpt-oss-120b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 62.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.1 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 58.7 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 60.6 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 64.4 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "80 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 58.2 }, "rule_version": 1 }, { "gpu_id": "h100-pcie-80gb", "llm_id": "glm-4.5-air", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "80 GB of memory is at least the 74.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 80, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.2 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "intel-arc-a310-lp-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "iris-xe-max-100-2gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "l4-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "l40-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "l40s-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k1200-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k2200-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k4200-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k600-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "quadro-k620-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m2000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m4000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m5000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-24gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "quadro-m6000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p1000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "5 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "5 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "5 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "5 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -22.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "5 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "5 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -45.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -57.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p2000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -69.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "5 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "5 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "5 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "5 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -22.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "5 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "5 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -45.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -57.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p2200-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "5 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 5, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -69.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p400-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p4000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p5000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p600-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p6000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "quadro-p620-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-4000-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-5000-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-6000-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "quadro-rtx-8000-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "quadro-t1000-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r7-240-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r7-250-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r7-260x-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r9-270-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r9-270x-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.5 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "3 GB of memory is below the 3.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.2 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -24.5 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "3 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.7 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.9 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.0 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -47.7 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -59.9 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.3 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.4 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.8 }, "rule_version": 1 }, { "gpu_id": "r9-280-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -71.8 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "3 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.5 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "3 GB of memory is below the 3.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.2 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -24.5 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "3 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.7 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.9 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.0 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -47.7 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -59.9 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.3 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.4 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.8 }, "rule_version": 1 }, { "gpu_id": "r9-280x-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "3 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 3, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -71.8 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "r9-290-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "r9-290x-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "r9-380-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "r9-390-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "r9-fury-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "r9-fury-x-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "r9-nano-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "radeon-520-2gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-0.6b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "1 GB of memory is below the 1.9 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.9 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 1.03 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.2 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.5 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.7 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.9 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.0 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -49.7 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -61.9 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.3 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.4 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.6 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -20.8 }, "rule_version": 1 }, { "gpu_id": "radeon-540-1gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "1 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 1, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -73.8 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "radeon-550-2gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6400-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6600-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.8 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.5 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.3 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.1 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.0 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 22.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.9 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.7 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.6 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.4 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.2 }, "rule_version": 1 }, { "gpu_id": "radeon-pro-w6800-32gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.8 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "radeon-r5-430-2gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "radeon-r7-450-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "radeon-rx-640-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "radeon-vii-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2000-ada-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-12gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2060-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2070-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "11 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "11 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "11 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.5 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "11 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "11 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.0 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "11 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -39.7 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -51.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.3 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "11 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.8 }, "rule_version": 1 }, { "gpu_id": "rtx-2080-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "11 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 11, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -63.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-lp-gigabyte", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-6gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3050-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-12gb-palit-infinity2-oc", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3060-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-8gb", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3070-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "gemma-3-12b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 9.1 GB required at a context length of 4,096 tokens, but not the 10.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -40.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -52.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -64.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 9.1 GB required at a context length of 4,096 tokens, but not the 10.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -40.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -52.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-10gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -64.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-12gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3080ti-12gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3090-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-3090ti-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.5 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "20 GB of memory covers the 18.7 GB required at a context length of 4,096 tokens, but not the 21.3 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-ada-20gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.5 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "20 GB of memory covers the 18.7 GB required at a context length of 4,096 tokens, but not the 21.3 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4000-sff-ada-20gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-lp-gigabyte", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-16gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4060-ti-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4070-ti-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4080-super-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-4090-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-4500-ada-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 22.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.2 }, "rule_version": 1 }, { "gpu_id": "rtx-5000-ada-32gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5050-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-16gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-msi-ventus2x", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5060-ti-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5070-ti-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5080-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.8 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.5 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.3 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.1 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 22.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.9 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.7 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.6 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.4 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.2 }, "rule_version": 1 }, { "gpu_id": "rtx-5090-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.8 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "rtx-6000-ada-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a1000", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-12gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a2000-6gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a400", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4000-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.5 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "20 GB of memory covers the 18.7 GB required at a context length of 4,096 tokens, but not the 21.3 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a4500-20gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-a5000-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-a5500-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 42.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 40.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.8 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 20.5 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 43.3 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 37.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.0 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 38.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "llama-3.1-70b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "48 GB of memory covers the 41.9 GB required at a context length of 4,096 tokens, but not the 50.7 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.1 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.9 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 28.6 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "48 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 32.4 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "48 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.2 }, "rule_version": 1 }, { "gpu_id": "rtx-a6000-48gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "48 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 48, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4000-blackwell-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.5 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.3 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.0 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 22.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.7 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.9 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.7 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.2 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-4500-blackwell-32gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 91.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 90.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 88.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 85.8 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 81.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 68.5 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 91.3 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 85.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 76.0 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 86.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "llama-3.1-70b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 50.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 45.3 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 81.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "gpt-oss-120b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 62.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 33.1 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 74.7 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 76.6 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 80.4 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "96 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 74.2 }, "rule_version": 1 }, { "gpu_id": "rtx-pro-6000-blackwell-96gb", "llm_id": "glm-4.5-air", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "96 GB of memory is at least the 74.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 96, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.2 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "rx-460-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-470-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-480-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-550-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-4gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-5500-xt-8gb-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-550x-4gb", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-560-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "6 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.3 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.2 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.4 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -21.5 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "6 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.9 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -14.0 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "6 GB of memory is below the 6.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -44.7 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.6 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -56.9 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.3 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.4 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.6 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.8 }, "rule_version": 1 }, { "gpu_id": "rx-5600-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "6 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 6, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -68.8 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-570-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-5700-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-5700-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-580-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-590-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-6400-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-6400-lp-kuroutoshikou", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "rx-6500-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-0.6b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-1.7b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-14b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "gemma-3-4b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "gemma-3-12b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "gemma-3-27b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "llama-3.1-8b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "llama-3.1-70b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "gpt-oss-20b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "gpt-oss-120b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "qwen3-30b-a3b-2507", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "mistral-small-3.2-24b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "phi-4", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "exaone-4.0-32b", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given.", "confidence": [ "B" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6500xt", "llm_id": "glm-4.5-air", "verdict": "unverified", "provisional": false, "rule_id": "r0", "reason": "The memory size of this graphics option is not stated in the cited specification, so no verdict is given. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": null, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": null }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-6600-le-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-6600-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-6600-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-6650-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.7 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.5 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "10 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "gemma-3-12b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "10 GB of memory covers the 9.1 GB required at a context length of 4,096 tokens, but not the 10.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.9 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.0 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "10 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -40.7 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.6 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -52.9 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "10 GB of memory is below the 10.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.8 }, "rule_version": 1 }, { "gpu_id": "rx-6700-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "10 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 10, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -64.8 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rx-6700-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rx-6750-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6800-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6900xt-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-6950-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-7600-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7600-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "rx-7700-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7800-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-gre-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.6 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.5 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.5 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 15.3 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.0 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.7 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "20 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.9 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "20 GB of memory covers the 18.7 GB required at a context length of 4,096 tokens, but not the 21.3 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.3 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.6 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "20 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "20 GB of memory is below the 20.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -1.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "20 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 20, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "rx-7900-xtx-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-9050-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9060-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "rx-9070-xt-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-vega-56-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "rx-vega-64-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "t1000-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "t400-4gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "2 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-1.7b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "2 GB of memory is below the 2.5 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -3.6 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.33 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.9 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.2 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.4 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -25.5 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 2.94 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.7 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.9 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.0 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -48.7 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.6 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -60.9 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.3 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.4 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.6 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.8 }, "rule_version": 1 }, { "gpu_id": "t400-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "2 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 2, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -72.8 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 1.9 GB required at a context length of 4,096 tokens, but not the 4.9 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.1 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "4 GB of memory covers the 2.5 GB required at a context length of 4,096 tokens, but not the 5.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.5 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-4b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "4 GB of memory covers the 3.9 GB required at a context length of 4,096 tokens, but not the 7.9 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-8b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.68 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.2 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.4 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -23.5 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "gemma-3-4b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3", "reason": "4 GB of memory is below the 4.2 GB required at a context length of 4,096 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -0.7 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 7.52 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.9 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -16.0 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "llama-3.1-8b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 4.58 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.6 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.7 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -10.6 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.9 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.3 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.4 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.6 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -17.8 }, "rule_version": 1 }, { "gpu_id": "t600-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "4 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 4, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -70.8 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "tesla-m40-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p100-pcie-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.3 GB required at a context length of 4,096 tokens, but not the 10.2 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.7 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-14b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.38 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -19.5 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "8 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.3 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "gemma-3-12b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3", "reason": "8 GB of memory is below the 9.1 GB required at a context length of 4,096 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.9 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -12.0 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "llama-3.1-8b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "8 GB of memory covers the 6.1 GB required at a context length of 4,096 tokens, but not the 9.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.7 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -6.6 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -54.9 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.3 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "phi-4", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 8.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.6 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -13.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p4-8gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "8 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 8, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -66.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "tesla-p40-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-t4-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-16gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 26.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 24.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.8 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.6 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 27.5 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.5 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 27.3 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 21.1 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.0 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 22.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -18.7 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 17.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -30.9 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.7 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 12.6 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "32 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.4 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "32 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.2 }, "rule_version": 1 }, { "gpu_id": "tesla-v100-pcie-32gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "32 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 32, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -42.8 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.1 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 18.4 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 16.1 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.8 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.6 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-32b", "verdict": "fits_short_context_only", "provisional": false, "rule_id": "r2", "reason": "24 GB of memory covers the 20.5 GB required at a context length of 4,096 tokens, but not the 27.5 GB required at 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 3.5 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 19.3 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 13.1 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "gemma-3-27b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 20.0 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.0 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 14.4 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -26.7 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 9.4 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": false, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.9 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.3 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.7 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 19.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.6 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "24 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.4 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "exaone-4.0-32b", "verdict": "fits_in_vram", "provisional": false, "rule_id": "r1", "reason": "24 GB of memory is at least the 21.8 GB required at a context length of 32,768 tokens.", "confidence": [ "B" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.2 }, "rule_version": 1 }, { "gpu_id": "titan-rtx-24gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "24 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: the size of the model file that is distributed is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 24, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.8 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "titan-v-12gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.1 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 4.1 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.8 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-14b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 14.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -15.5 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 7.3 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.1 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -8.0 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "12 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed and this model's layer count and this model's key/value head count and this model's head dimension are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 2.4 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -38.7 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "gpt-oss-20b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 12.82 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -2.6 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -50.9 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.3 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "mistral-small-3.2-24b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 13.35 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -7.4 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "phi-4", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "12 GB of memory covers the 10.1 GB required at a context length of 4,096 tokens, but not the 15.6 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.9 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -9.8 }, "rule_version": 1 }, { "gpu_id": "titan-xp-12gb", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "12 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size and the size of the model file that is distributed are recorded here from third-party or community material, so grade D is shown alongside B.", "confidence": [ "B", "D" ], "basis": { "vram_gb": 12, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -62.8 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-0.6b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 0.37, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 1.9, "required_working_context_gib": 4.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.1 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-1.7b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 5.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 1.03, "kv_cache_short_context_gib": 0.44, "kv_cache_working_context_gib": 3.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 2.5, "required_working_context_gib": 5.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 10.4 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 7.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 2.33, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 3.9, "required_working_context_gib": 7.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 8.1 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.2 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.68, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 4.5, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.3, "required_working_context_gib": 10.2, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.8 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-14b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.4 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.38, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 14.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.6 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.40 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.4, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 8.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.5, "required_working_context_gib": 27.5, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -11.5 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "gemma-3-4b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 4.7 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 2.94, "kv_cache_short_context_gib": 0.19, "kv_cache_working_context_gib": 0.74, "runtime_overhead_gib": 1.0, "required_short_context_gib": 4.2, "required_working_context_gib": 4.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 11.3 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "gemma-3-12b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 10.9 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 7.52, "kv_cache_short_context_gib": 0.56, "kv_cache_working_context_gib": 2.31, "runtime_overhead_gib": 1.0, "required_short_context_gib": 9.1, "required_working_context_gib": 10.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 5.1 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "gemma-3-27b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 16.05 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_0 (QAT)", "weights_gib": 16.05, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 2.91, "runtime_overhead_gib": 1.0, "required_short_context_gib": 17.8, "required_working_context_gib": 20.0, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -4.0 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "llama-3.1-8b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 9.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 4.58, "kv_cache_short_context_gib": 0.5, "kv_cache_working_context_gib": 4.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 6.1, "required_working_context_gib": 9.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 6.4 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "llama-3.1-70b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 39.60 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 39.6, "kv_cache_short_context_gib": 1.25, "kv_cache_working_context_gib": 10.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 41.9, "required_working_context_gib": 50.7, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -34.7 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "gpt-oss-20b", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 14.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 12.82, "kv_cache_short_context_gib": 0.1, "kv_cache_working_context_gib": 0.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 14.0, "required_working_context_gib": 14.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.4 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "gpt-oss-120b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 60.77 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "MXFP4 (the vendor's own native quantisation)", "weights_gib": 60.77, "kv_cache_short_context_gib": 0.15, "kv_cache_working_context_gib": 1.13, "runtime_overhead_gib": 1.0, "required_short_context_gib": 62.0, "required_working_context_gib": 62.9, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -46.9 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "qwen3-30b-a3b-2507", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 17.28 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 17.28, "kv_cache_short_context_gib": 0.38, "kv_cache_working_context_gib": 3.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 18.7, "required_working_context_gib": 21.3, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.3 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "mistral-small-3.2-24b", "verdict": "fits_short_context_only", "provisional": true, "rule_id": "r2", "reason": "16 GB of memory covers the 15.0 GB required at a context length of 4,096 tokens, but not the 19.4 GB required at 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 13.35, "kv_cache_short_context_gib": 0.62, "kv_cache_working_context_gib": 5.0, "runtime_overhead_gib": 1.0, "required_short_context_gib": 15.0, "required_working_context_gib": 19.4, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 1.0 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "phi-4", "verdict": "fits_in_vram", "provisional": true, "rule_id": "r1", "reason": "16 GB of memory is at least the 15.6 GB required at a context length of 32,768 tokens. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 8.28, "kv_cache_short_context_gib": 0.78, "kv_cache_working_context_gib": 6.25, "runtime_overhead_gib": 1.0, "required_short_context_gib": 10.1, "required_working_context_gib": 15.6, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": 0.4 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "exaone-4.0-32b", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 18.02 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 18.02, "kv_cache_short_context_gib": 1.0, "kv_cache_working_context_gib": 2.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 20.1, "required_working_context_gib": 21.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -5.8 }, "rule_version": 1 }, { "gpu_id": "vega-frontier-edition-ref", "llm_id": "glm-4.5-air", "verdict": "exceeds_vram", "provisional": true, "rule_id": "r3b", "reason": "16 GB of memory is smaller than the 67.96 GB model file itself, so shortening the context does not make it fit in memory. This outcome is no stronger than the figures it was calculated from: this card's memory size is recorded here from material whose source the record does not identify, so grade E is shown alongside B.", "confidence": [ "B", "E" ], "basis": { "vram_gb": 16, "quant": "Q4_K_M", "weights_gib": 67.96, "kv_cache_short_context_gib": 0.72, "kv_cache_working_context_gib": 5.75, "runtime_overhead_gib": 1.0, "required_short_context_gib": 69.7, "required_working_context_gib": 74.8, "short_context_tokens": 4096, "working_context_tokens": 32768, "kv_cache_included": true, "headroom_gib": -58.8 }, "rule_version": 1 } ]