{
 "as_of": "2026-10-12",
 "license": "CC-BY-4.0",
 "license_url": "https://creativecommons.org/licenses/by/4.0/",
 "cite_as": "The Local Bench (thelocalbench.com), data as of 2026-10-12, CC BY 4.0",
 "source_url": "https://thelocalbench.com",
 "models": [
  {
   "id": "llama-3.1-8b",
   "name": "Llama 3.1 8B Instruct",
   "org": "Meta",
   "task": "general",
   "hf_repo": "meta-llama/Llama-3.1-8B-Instruct",
   "gguf_repo": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
   "license": {
    "value": "Llama 3.1 Community License",
    "flag": "verified",
    "source": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/blob/main/LICENSE"
   },
   "datasheet": "TLB-M-0001",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The default small model for a reason: well supported by every runtime, a 128K window, and it fits an 8 GB card at Q4.",
   "arch": {
    "params_total": {
     "value": 8.03,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/Llama-3.1-8B-Instruct",
     "note": "safetensors.total from the HF API. Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "params_active": {
     "value": 8.03,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/Llama-3.1-8B-Instruct",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "q_heads": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "hidden_size": {
     "value": 4096,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json"
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 4.02,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q3_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 4.92,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 5.73,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 6.6,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 8.54,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 16.06,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 16.0 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in bartowski/Meta-Llama-3.1-8B-Instruct-GGUF."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1186.6,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'llama-3.1-8b-instruct': 1186.6 (95% CI 1182.5–1190.7, 49,605 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "qwen3-8b",
   "name": "Qwen3 8B",
   "org": "Alibaba Qwen",
   "task": "general",
   "hf_repo": "Qwen/Qwen3-8B",
   "gguf_repo": "Qwen/Qwen3-8B-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0002",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "A small dense model with switchable thinking. The pick for 8 to 12 GB cards when you want Apache-2.0 terms.",
   "arch": {
    "params_total": {
     "value": 8.19,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-8B",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 8.19,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-8B",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 36,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json"
    },
    "q_heads": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json"
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json"
    },
    "hidden_size": {
     "value": 4096,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json"
    },
    "context": {
     "value": 32768,
     "flag": "sourced",
     "unit": "tokens",
     "source": "https://huggingface.co/Qwen/Qwen3-8B",
     "note": "config.json lists max_position_embeddings 40,960; the model card gives 32,768 native and 131,072 with YaRN."
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 4,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 3.91 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-8B-GGUF."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 5.03,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-8B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 5.85,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-8B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 6.73,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-8B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 8.71,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-8B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 16.38,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 16.0 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-8B-GGUF."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1270,
     "flag": "sample",
     "note": "Placeholder: this model has no entry in the LMArena text leaderboard dataset (checked 2026-10-12). Not a real score."
    }
   }
  },
  {
   "id": "qwen3-14b",
   "name": "Qwen3 14B",
   "org": "Alibaba Qwen",
   "task": "general",
   "hf_repo": "Qwen/Qwen3-14B",
   "gguf_repo": "Qwen/Qwen3-14B-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0003",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The sweet spot for 16 GB cards: Q4_K_M weights are 9 GB, which leaves room for a long KV cache.",
   "arch": {
    "params_total": {
     "value": 14.77,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-14B",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 14.77,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-14B",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 40,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json"
    },
    "q_heads": {
     "value": 40,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json"
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json"
    },
    "hidden_size": {
     "value": 5120,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json"
    },
    "context": {
     "value": 32768,
     "flag": "sourced",
     "unit": "tokens",
     "source": "https://huggingface.co/Qwen/Qwen3-14B",
     "note": "config.json lists max_position_embeddings 40,960; the model card gives 32,768 native and 131,072 with YaRN."
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 7.22,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 3.91 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-14B-GGUF."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 9,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-14B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 10.51,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-14B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 12.12,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-14B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 15.7,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-14B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 29.54,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 16.0 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-14B-GGUF."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1301,
     "flag": "sample",
     "note": "Placeholder: this model has no entry in the LMArena text leaderboard dataset (checked 2026-10-12). Not a real score."
    }
   }
  },
  {
   "id": "gpt-oss-20b",
   "name": "gpt-oss-20b",
   "org": "OpenAI",
   "task": "general",
   "hf_repo": "openai/gpt-oss-20b",
   "gguf_repo": "ggml-org/gpt-oss-20b-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0004",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "A mixture-of-experts model shipped natively in MXFP4. Only 3.6B parameters are read per token, so it is fast on almost anything with 16 GB.",
   "arch": {
    "params_total": {
     "value": 20.91,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/openai/gpt-oss-20b",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 3.6,
     "flag": "sourced",
     "unit": "B",
     "source": "https://huggingface.co/openai/gpt-oss-20b",
     "note": "Active parameters per token from the model card."
    },
    "layers": {
     "value": 24,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "q_heads": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "head_dim": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "hidden_size": {
     "value": 2880,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    },
    "sliding_window": {
     "value": 128,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json",
     "note": "Local attention window."
    },
    "sliding_layers": {
     "value": 12,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json",
     "note": "Layers using sliding-window attention (from layer_types / sliding_window_pattern)."
    },
    "experts": {
     "value": "32 experts, 4 active per token",
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"
    }
   },
   "quants": {
    "MXFP4": {
     "bpw": 4.25,
     "file_gb": {
      "value": 12.11,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/ggml-org/gpt-oss-20b-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (MXFP4) from the HF API tree listing."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1287.4,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'gpt-oss-20b': 1287.4 (95% CI 1281.0–1293.7, 10,418 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "mistral-small-3.2-24b",
   "name": "Mistral Small 3.2 24B",
   "org": "Mistral AI",
   "task": "general",
   "hf_repo": "mistralai/Mistral-Small-3.2-24B-Instruct-2506",
   "gguf_repo": "bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0005",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "A dense 24B with vision input and a 128K window. Q4_K_M is 14.3 GB, a comfortable fit on a 24 GB card at long context.",
   "arch": {
    "params_total": {
     "value": 24.01,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/mistralai/Mistral-Small-3.2-24B-Instruct-2506",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 24.01,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/mistralai/Mistral-Small-3.2-24B-Instruct-2506",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 40,
     "flag": "verified",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    },
    "q_heads": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    },
    "hidden_size": {
     "value": 5120,
     "flag": "verified",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json"
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 11.47,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q3_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 14.33,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 16.76,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 19.35,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 25.05,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 47.15,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (BF16) from the HF API tree listing."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1338.1,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'mistral-small-2506': 1338.1 (95% CI 1332.9–1343.3, 17,466 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "gemma-3-27b",
   "name": "Gemma 3 27B IT",
   "org": "Google DeepMind",
   "task": "general",
   "hf_repo": "google/gemma-3-27b-it",
   "gguf_repo": "bartowski/google_gemma-3-27b-it-GGUF",
   "license": {
    "value": "Gemma Terms of Use",
    "flag": "verified",
    "source": "https://ai.google.dev/gemma/terms"
   },
   "datasheet": "TLB-M-0006",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The 24 GB-card workhorse. Five of every six layers only look at the last 1,024 tokens, so the KV cache stays small as context grows.",
   "arch": {
    "params_total": {
     "value": 27.43,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/gemma-3-27b-it",
     "note": "safetensors.total from the HF API. Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "params_active": {
     "value": 27.43,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/gemma-3-27b-it",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 62,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "kv_heads": {
     "value": 16,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "q_heads": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "hidden_size": {
     "value": 5376,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json"
    },
    "sliding_window": {
     "value": 1024,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Local attention window."
    },
    "sliding_layers": {
     "value": 52,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json",
     "note": "Layers using sliding-window attention (from layer_types / sliding_window_pattern)."
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 13.44,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q3_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 16.55,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 19.27,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 22.17,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 28.71,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 54.03,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (BF16) from the HF API tree listing."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1357.7,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'gemma-3-27b-it': 1357.7 (95% CI 1354.1–1361.4, 46,672 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "qwen3-30b-a3b",
   "name": "Qwen3 30B-A3B",
   "org": "Alibaba Qwen",
   "task": "general",
   "hf_repo": "Qwen/Qwen3-30B-A3B",
   "gguf_repo": "Qwen/Qwen3-30B-A3B-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0007",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "A mixture-of-experts model that reads only 3.3B parameters per token. Needs 24 GB to hold, but decodes like a small model.",
   "arch": {
    "params_total": {
     "value": 30.53,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 3.3,
     "flag": "sourced",
     "unit": "B",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B",
     "note": "Active parameters per token from the model card."
    },
    "layers": {
     "value": 48,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    },
    "kv_heads": {
     "value": 4,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    },
    "q_heads": {
     "value": 32,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    },
    "hidden_size": {
     "value": 2048,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    },
    "context": {
     "value": 32768,
     "flag": "sourced",
     "unit": "tokens",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B",
     "note": "config.json lists max_position_embeddings 40,960; the model card gives 32,768 native and 131,072 with YaRN."
    },
    "experts": {
     "value": "128 experts, 8 active per token",
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json"
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 14.92,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 3.91 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-30B-A3B-GGUF."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 18.56,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 21.73,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 25.09,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 32.48,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 61.06,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 16.0 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-30B-A3B-GGUF."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1316.8,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'qwen3-30b-a3b': 1316.8 (95% CI 1312.1–1321.5, 26,160 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "qwen3-32b",
   "name": "Qwen3 32B",
   "org": "Alibaba Qwen",
   "task": "general",
   "hf_repo": "Qwen/Qwen3-32B",
   "gguf_repo": "Qwen/Qwen3-32B-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0008",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The largest dense Qwen3. Q4_K_M weights are 19.8 GB, which is tight on a 24 GB card once context grows.",
   "arch": {
    "params_total": {
     "value": 32.76,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-32B",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 32.76,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/Qwen/Qwen3-32B",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    "q_heads": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    "hidden_size": {
     "value": 5120,
     "flag": "verified",
     "source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    "context": {
     "value": 32768,
     "flag": "sourced",
     "unit": "tokens",
     "source": "https://huggingface.co/Qwen/Qwen3-32B",
     "note": "config.json lists max_position_embeddings 40,960; the model card gives 32,768 native and 131,072 with YaRN."
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 16.01,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 3.91 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-32B-GGUF."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 19.76,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-32B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 23.21,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-32B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 26.88,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-32B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 34.82,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/Qwen/Qwen3-32B-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "F16": {
     "bpw": 16,
     "file_gb": {
      "value": 65.52,
      "flag": "estimate",
      "unit": "GB",
      "note": "params_total × 16.0 bpw ÷ 8 (brief §7.3 fallback); no GGUF of this quant in Qwen/Qwen3-32B-GGUF."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1340.1,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'qwen3-32b': 1340.1 (95% CI 1330.8–1349.4, 3,926 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "llama-3.3-70b",
   "name": "Llama 3.3 70B Instruct",
   "org": "Meta",
   "task": "general",
   "hf_repo": "meta-llama/Llama-3.3-70B-Instruct",
   "gguf_repo": "bartowski/Llama-3.3-70B-Instruct-GGUF",
   "license": {
    "value": "Llama 3.3 Community License",
    "flag": "verified",
    "source": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/blob/main/LICENSE"
   },
   "datasheet": "TLB-M-0009",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The classic 48 GB model. Two used 3090s or a large unified-memory box; bandwidth, not capacity, sets the speed.",
   "arch": {
    "params_total": {
     "value": 70.55,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/Llama-3.3-70B-Instruct",
     "note": "safetensors.total from the HF API. Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "params_active": {
     "value": 70.55,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/unsloth/Llama-3.3-70B-Instruct",
     "note": "Dense: all parameters active."
    },
    "layers": {
     "value": 80,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "q_heads": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "head_dim": {
     "value": 128,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "hidden_size": {
     "value": 8192,
     "flag": "verified",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json",
     "note": "Read from an ungated mirror of the gated official repo; identical architecture."
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json"
    }
   },
   "quants": {
    "Q3_K_M": {
     "bpw": 3.91,
     "file_gb": {
      "value": 34.27,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q3_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.071,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q4_K_M": {
     "bpw": 4.85,
     "file_gb": {
      "value": 42.52,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.017,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q5_K_M": {
     "bpw": 5.69,
     "file_gb": {
      "value": 49.95,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."
     },
     "kld": {
      "value": 0.008,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q6_K": {
     "bpw": 6.56,
     "file_gb": {
      "value": 57.89,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q6_K) from the HF API tree listing."
     },
     "kld": {
      "value": 0.004,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    },
    "Q8_0": {
     "bpw": 8.5,
     "file_gb": {
      "value": 74.98,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (Q8_0) from the HF API tree listing."
     },
     "kld": {
      "value": 0.001,
      "flag": "sample",
      "note": "Placeholder until KLD runs (brief §4.2)."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1274.2,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'llama-3.3-70b-instruct': 1274.2 (95% CI 1270.8–1277.7, 54,504 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  },
  {
   "id": "gpt-oss-120b",
   "name": "gpt-oss-120b",
   "org": "OpenAI",
   "task": "general",
   "hf_repo": "openai/gpt-oss-120b",
   "gguf_repo": "ggml-org/gpt-oss-120b-GGUF",
   "license": {
    "value": "Apache-2.0",
    "flag": "verified",
    "source": "https://www.apache.org/licenses/LICENSE-2.0"
   },
   "datasheet": "TLB-M-0010",
   "rev": "2026.10",
   "as_of": "2026-10-12",
   "summary": "The model that sells 128 GB unified-memory boxes: 63 GB of MXFP4 weights, but only 5.1B parameters read per token.",
   "arch": {
    "params_total": {
     "value": 116.83,
     "flag": "verified",
     "unit": "B",
     "source": "https://huggingface.co/api/models/openai/gpt-oss-120b",
     "note": "safetensors.total from the HF API."
    },
    "params_active": {
     "value": 5.1,
     "flag": "sourced",
     "unit": "B",
     "source": "https://huggingface.co/openai/gpt-oss-120b",
     "note": "Active parameters per token from the model card."
    },
    "layers": {
     "value": 36,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "kv_heads": {
     "value": 8,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "q_heads": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "head_dim": {
     "value": 64,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "hidden_size": {
     "value": 2880,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "context": {
     "value": 131072,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    },
    "sliding_window": {
     "value": 128,
     "flag": "verified",
     "unit": "tokens",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json",
     "note": "Local attention window."
    },
    "sliding_layers": {
     "value": 18,
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json",
     "note": "Layers using sliding-window attention (from layer_types / sliding_window_pattern)."
    },
    "experts": {
     "value": "128 experts, 4 active per token",
     "flag": "verified",
     "source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json"
    }
   },
   "quants": {
    "MXFP4": {
     "bpw": 4.25,
     "file_gb": {
      "value": 63.39,
      "flag": "verified",
      "unit": "GB",
      "source": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF/tree/main",
      "note": "Sum of GGUF file sizes (MXFP4) from the HF API tree listing."
     }
    }
   },
   "quality": {
    "arena_text": {
     "value": 1365.4,
     "flag": "sourced",
     "source": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
     "note": "LMArena text, overall, entry 'gpt-oss-120b': 1365.4 (95% CI 1361.1–1369.8, 30,218 votes), leaderboard published 2026-10-08. CC-BY-4.0."
    }
   }
  }
 ],
 "hardware": [
  {
   "id": "rtx-5060-ti-16g",
   "name": "GeForce RTX 5060 Ti 16 GB",
   "vendor": "NVIDIA",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "GDDR7 lifts bandwidth by about half over the 4060 Ti for a similar price.",
   "memory_gb": {
    "value": 16,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/",
    "note": "16 GB variant, nvidia.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 448,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_50_series",
    "note": "NVIDIA's GeForce spec pages list bus width but not bandwidth; figure from the cited Wikipedia spec table (bus width cross-checked on nvidia.com, 2026-10-12)."
   },
   "fp16_tflops": {
    "value": 47,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 32,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 250,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 180,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/",
    "note": "Total Graphics Power (16 GB variant), nvidia.com."
   },
   "price_usd": {
    "value": 779.99,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.newegg.com/p/pl?d=rtx+5060+ti+16gb&N=4131",
    "note": "Lowest new in-stock 16 GB listing on Newegg US (MSI Shadow). Checked 2026-10-12."
   },
   "price_cad": {
    "value": 1119.99,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.newegg.ca/p/pl?d=rtx+5060+ti+16gb&N=4131",
    "note": "Lowest new in-stock 16 GB listing on Newegg Canada (ASUS Dual). Checked 2026-10-12."
   }
  },
  {
   "id": "rtx-4060-ti-16g",
   "name": "GeForce RTX 4060 Ti 16 GB",
   "vendor": "NVIDIA",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "Cheapest new 16 GB card. Narrow 128-bit bus caps decode speed.",
   "memory_gb": {
    "value": 16,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4060-4060ti/",
    "note": "16 GB variant, nvidia.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 288,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
    "note": "NVIDIA's GeForce spec pages list bus width but not bandwidth; figure from the cited Wikipedia spec table (bus width cross-checked on nvidia.com, 2026-10-12)."
   },
   "fp16_tflops": {
    "value": 44,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 32,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 230,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 165,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4060-4060ti/",
    "note": "Total Graphics Power (16 GB variant), nvidia.com."
   },
   "price_usd": {
    "value": 844.37,
    "flag": "estimate",
    "unit": "USD",
    "source": "https://www.newegg.ca/p/pl?d=rtx+4060+ti+16gb&N=4131",
    "note": "No valid new 16 GB listing found on Newegg US; Newegg Canada price ÷ 1.4271 (Bank of Canada)."
   },
   "price_cad": {
    "value": 1205,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.newegg.ca/p/pl?d=rtx+4060+ti+16gb&N=4131",
    "note": "Lowest new in-stock 16 GB listing on Newegg Canada (MSI Gaming). Checked 2026-10-12."
   }
  },
  {
   "id": "rx-7900-xtx-24g",
   "name": "Radeon RX 7900 XTX 24 GB",
   "vendor": "AMD",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "24 GB for less than a used 3090 at some retailers; ROCm and Vulkan backends both work in llama.cpp.",
   "memory_gb": {
    "value": 24,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 960,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "fp16_tflops": {
    "value": 123,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 64,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 400,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 355,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xtx.html",
    "note": "Typical Board Power (desktop), amd.com."
   },
   "price_usd": {
    "value": 899,
    "flag": "sample",
    "unit": "USD",
    "note": "Sample price."
   },
   "price_cad": {
    "value": 1282.96,
    "flag": "sample",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Converted from a sample USD value."
   }
  },
  {
   "id": "rtx-3090-24g-used",
   "name": "GeForce RTX 3090 24 GB (used)",
   "vendor": "NVIDIA",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "The reference card of the hobby: 24 GB and 936 GB/s for about the price of a new mid-range GPU.",
   "memory_gb": {
    "value": 24,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 936,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_30_series",
    "note": "NVIDIA's GeForce spec pages list bus width but not bandwidth; figure from the cited Wikipedia spec table (bus width cross-checked on nvidia.com, 2026-10-12)."
   },
   "fp16_tflops": {
    "value": 71,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 64,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 420,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 350,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/",
    "note": "Graphics Card Power (Founders Edition), nvidia.com."
   },
   "price_usd": {
    "value": 1200,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.getpcparts.com/market-prices/gpu-graphics-cards/rtx-3090",
    "note": "eBay sold-listing median ≈ US$1,199–1,206, Oct 2026, via the getpcparts aggregator (eBay itself blocks automated access). Checked 2026-10-12."
   },
   "price_cad": {
    "value": 1712.52,
    "flag": "estimate",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured."
   }
  },
  {
   "id": "cpu-ryzen-9-9950x-128g",
   "name": "CPU only · Ryzen 9 9950X + 128 GB DDR5",
   "vendor": "AMD",
   "class": "CPU reference",
   "kind": "cpu",
   "as_of": "2026-10-12",
   "summary": "The no-GPU baseline: dual-channel DDR5 at 89.6 GB/s. Fine for MoE, slow for dense.",
   "memory_gb": {
    "value": 128,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html",
    "note": "Build configuration (2 × 64 GB); amd.com lists 256 GB max, 2 channels."
   },
   "bandwidth_gbs": {
    "value": 89.6,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html",
    "note": "2 channels × DDR5-5600 (amd.com max supported speed for 2 DIMMs) × 8 bytes = 89.6 GB/s (derived from vendor figures)."
   },
   "fp16_tflops": {
    "value": 2.5,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 89.6,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html",
    "note": "Same as system memory."
   },
   "load_watts": {
    "value": 200,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 170,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html",
    "note": "Default TDP, amd.com."
   },
   "price_usd": {
    "value": 1400,
    "flag": "sample",
    "unit": "USD",
    "note": "Sample price."
   },
   "price_cad": {
    "value": 1997.94,
    "flag": "sample",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Converted from a sample USD value."
   }
  },
  {
   "id": "2x-rtx-3090-48g-used",
   "name": "2× RTX 3090 (used) 48 GB",
   "vendor": "NVIDIA",
   "class": "Multi-GPU",
   "kind": "multi",
   "as_of": "2026-10-12",
   "summary": "Layer split adds capacity, not bandwidth, for a single stream (brief §7.3).",
   "memory_gb": {
    "value": 48,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/",
    "note": "2 × 24 GB (nvidia.com)."
   },
   "bandwidth_gbs": {
    "value": 936,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_30_series",
    "note": "Per card; layer split does not add bandwidth for a single stream."
   },
   "fp16_tflops": {
    "value": 71,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 128,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 780,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 700,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/",
    "note": "2 × 350 W Graphics Card Power, nvidia.com."
   },
   "price_usd": {
    "value": 2400,
    "flag": "estimate",
    "unit": "USD",
    "source": "https://www.getpcparts.com/market-prices/gpu-graphics-cards/rtx-3090",
    "note": "2 × the eBay-sold median for a used RTX 3090 (≈US$1,200, Oct 2026); excludes the second-PSU/board upgrade."
   },
   "gpus": 2,
   "price_cad": {
    "value": 3425.04,
    "flag": "estimate",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured."
   }
  },
  {
   "id": "mac-studio-m5-max",
   "name": "Mac Studio M5 Max (base)",
   "vendor": "Apple",
   "class": "Apple unified",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "Quiet and small. The base memory size limits it to ~30B dense models at Q4.",
   "memory_gb": {
    "value": 36,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Base configuration, apple.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 460,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Base M5 Max (32-core GPU) is 460 GB/s; the 40-core GPU option is 614 GB/s. apple.com, checked 2026-10-12. Corrects the mockup value of 614."
   },
   "fp16_tflops": {
    "value": 40,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 460,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Unified memory."
   },
   "load_watts": {
    "value": 150,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 480,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Maximum continuous power of the whole Mac Studio (apple.com); Apple publishes no chip TDP."
   },
   "price_usd": {
    "value": 2499,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.apple.com/shop/buy-mac/mac-studio",
    "note": "Apple US store, base configuration. Checked 2026-10-12."
   },
   "apple": true,
   "price_cad": {
    "value": 3499,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.apple.com/ca/shop/buy-mac/mac-studio",
    "note": "Apple Canada store, base configuration. Checked 2026-10-12."
   }
  },
  {
   "id": "ryzen-ai-max-395-128g",
   "name": "Ryzen AI Max+ 395 mini PC 128 GB",
   "vendor": "AMD",
   "class": "Unified memory",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "128 GB of unified memory in a small box. Capacity for 100B-class MoE models; bandwidth limits dense models.",
   "memory_gb": {
    "value": 128,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html",
    "note": "Max memory 128 GB, amd.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 256,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html",
    "note": "256-bit LPDDR5x-8000 per amd.com → 256 × 8000 MT/s ÷ 8 = 256 GB/s (derived from vendor figures)."
   },
   "fp16_tflops": {
    "value": 59,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 256,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 120,
    "flag": "sourced",
    "unit": "W",
    "source": "brief §7.3",
    "note": "120 W load is the brief's break-even assumption."
   },
   "tdp_w": {
    "value": 120,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html",
    "note": "Top of the 45–120 W configurable TDP range (55 W default), amd.com."
   },
   "price_usd": {
    "value": 3499.99,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/",
    "note": "US$3,499.99 (CFG)."
   },
   "apple": false,
   "price_cad": {
    "value": 4994.84,
    "flag": "estimate",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured."
   }
  },
  {
   "id": "rtx-4090-24g",
   "name": "GeForce RTX 4090 24 GB",
   "vendor": "NVIDIA",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "Fastest 24 GB card; the extra compute shows up in prefill more than decode.",
   "memory_gb": {
    "value": 24,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 1008,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_40_series",
    "note": "NVIDIA's GeForce spec pages list bus width but not bandwidth; figure from the cited Wikipedia spec table (bus width cross-checked on nvidia.com, 2026-10-12)."
   },
   "fp16_tflops": {
    "value": 165,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 64,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 500,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 450,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/",
    "note": "Total Graphics Power, nvidia.com."
   },
   "price_usd": {
    "value": 4395,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.newegg.com/p/pl?d=rtx+4090+24gb+graphics+card&N=4131",
    "note": "Lowest new in-stock RTX 4090 listing on Newegg US (incl. marketplace sellers; card discontinued, scarce). Checked 2026-10-12."
   },
   "price_cad": {
    "value": 5599.99,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.newegg.ca/p/pl?d=rtx+4090+24gb+graphics+card&N=4131",
    "note": "Lowest new in-stock listing on Newegg Canada (MSI Gaming Trio, ships from Canada). Checked 2026-10-12."
   }
  },
  {
   "id": "dgx-spark-128g",
   "name": "NVIDIA DGX Spark FE 128 GB",
   "vendor": "NVIDIA",
   "class": "Unified memory",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "CUDA on 128 GB of unified memory. Similar bandwidth to Strix Halo, more compute for prefill.",
   "memory_gb": {
    "value": 128,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 273,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "fp16_tflops": {
    "value": 100,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 273,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 170,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 140,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/",
    "note": "GB10 TDP (CPU+GPU), nvidia.com; system power supply is 240 W."
   },
   "price_usd": {
    "value": 4699,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://vucense.com/tech-reviews/compute-chips/dgx-spark-vs-ryzen-ai-max-395-local-ai-workstation-2026/",
    "note": "US$4,699 (secondary source)."
   },
   "apple": false,
   "price_cad": {
    "value": 6705.94,
    "flag": "estimate",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured."
   }
  },
  {
   "id": "macbook-pro-m4-max-128g",
   "name": "MacBook Pro M4 Max 40c · 128 GB",
   "vendor": "Apple",
   "class": "Apple unified",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "A laptop with twice Strix Halo's bandwidth. MLX and Metal llama.cpp are both mature.",
   "memory_gb": {
    "value": 128,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://support.apple.com/en-us/121553",
    "note": "Configurable to 128 GB on the 16-core-CPU/40-core-GPU M4 Max, Apple tech specs, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 546,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://support.apple.com/en-us/121553",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "fp16_tflops": {
    "value": 34,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 546,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://support.apple.com/en-us/121553",
    "note": "Unified memory."
   },
   "load_watts": {
    "value": 140,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "price_usd": {
    "value": 4699,
    "flag": "sample",
    "unit": "USD",
    "note": "Sample price."
   },
   "apple": true,
   "price_cad": {
    "value": 6705.94,
    "flag": "sample",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Converted from a sample USD value."
   }
  },
  {
   "id": "mac-studio-m5-ultra",
   "name": "Mac Studio M5 Ultra (base)",
   "vendor": "Apple",
   "class": "Apple unified",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "1.2 TB/s of unified bandwidth: the fastest way to run 70B dense models at home.",
   "memory_gb": {
    "value": 96,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Base configuration, apple.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 1200,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "1.2 TB/s, apple.com, checked 2026-10-12."
   },
   "fp16_tflops": {
    "value": 80,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 1200,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Unified memory."
   },
   "load_watts": {
    "value": 240,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 480,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Maximum continuous power of the whole Mac Studio (apple.com); Apple publishes no chip TDP."
   },
   "price_usd": {
    "value": 5499,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.apple.com/shop/buy-mac/mac-studio",
    "note": "Apple US store, base M5 Ultra configuration. Checked 2026-10-12."
   },
   "apple": true,
   "price_cad": {
    "value": 7799,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.apple.com/ca/shop/buy-mac/mac-studio",
    "note": "Apple Canada store, base M5 Ultra configuration. Checked 2026-10-12."
   }
  },
  {
   "id": "rtx-5090-32g",
   "name": "GeForce RTX 5090 32 GB",
   "vendor": "NVIDIA",
   "class": "Discrete GPU",
   "kind": "gpu",
   "as_of": "2026-10-12",
   "summary": "32 GB and 1.8 TB/s: the fastest single consumer card for decode.",
   "memory_gb": {
    "value": 32,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 1792,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://en.wikipedia.org/wiki/GeForce_RTX_50_series",
    "note": "NVIDIA's GeForce spec pages list bus width but not bandwidth; figure from the cited Wikipedia spec table (bus width cross-checked on nvidia.com, 2026-10-12)."
   },
   "fp16_tflops": {
    "value": 209,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 64,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 64,
    "flag": "sample",
    "unit": "GB/s"
   },
   "load_watts": {
    "value": 650,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 575,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/",
    "note": "Total Graphics Power, nvidia.com."
   },
   "price_usd": {
    "value": 7199.99,
    "flag": "sourced",
    "unit": "USD",
    "source": "https://www.newegg.com/p/pl?d=rtx+5090+32gb+graphics+card&N=4131",
    "note": "Lowest new in-stock listing on Newegg US (PNY ARGB OC), far above the US$1,999 launch MSRP. Checked 2026-10-12."
   },
   "price_cad": {
    "value": 9036.35,
    "flag": "sourced",
    "unit": "CAD",
    "source": "https://www.newegg.ca/p/pl?d=rtx+5090+32gb+graphics+card&N=4131",
    "note": "Lowest new in-stock listing on Newegg Canada (ASUS TUF, marketplace seller). Checked 2026-10-12."
   }
  },
  {
   "id": "mac-studio-m5-ultra-512g",
   "name": "Mac Studio M5 Ultra · 512 GB",
   "vendor": "Apple",
   "class": "Apple unified",
   "kind": "unified",
   "as_of": "2026-10-12",
   "summary": "Room for frontier-scale open MoE models in one box.",
   "memory_gb": {
    "value": 512,
    "flag": "sourced",
    "unit": "GB",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Configurable to 512 GB (M5 Ultra 36-core CPU / 80-core GPU), apple.com, checked 2026-10-12."
   },
   "bandwidth_gbs": {
    "value": 1200,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Vendor spec page, checked 2026-10-12."
   },
   "fp16_tflops": {
    "value": 80,
    "flag": "sample",
    "unit": "TFLOPS",
    "note": "Sample FP16 dense compute; pending vendor-sheet citation."
   },
   "host_ram_gb": {
    "value": 0,
    "flag": "sample",
    "unit": "GB",
    "note": "Assumed host RAM for offload."
   },
   "host_bandwidth_gbs": {
    "value": 1200,
    "flag": "sourced",
    "unit": "GB/s",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Unified memory."
   },
   "load_watts": {
    "value": 260,
    "flag": "sample",
    "unit": "W",
    "note": "Sample wall power under load."
   },
   "tdp_w": {
    "value": 480,
    "flag": "sourced",
    "unit": "W",
    "source": "https://www.apple.com/mac-studio/specs/",
    "note": "Maximum continuous power of the whole Mac Studio (apple.com)."
   },
   "price_usd": {
    "value": 9999,
    "flag": "sample",
    "unit": "USD",
    "note": "Sample price."
   },
   "apple": true,
   "price_cad": {
    "value": 14269.57,
    "flag": "sample",
    "unit": "CAD",
    "source": "https://www.bankofcanada.ca/valet/observations/FXUSDCAD/json?recent=1",
    "note": "USD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Converted from a sample USD value."
   }
  }
 ],
 "builds": [
  {
   "id": "value-16gb",
   "name": "Value 16 GB",
   "label": "Best under $1,200",
   "hardware": "rtx-5060-ti-16g",
   "summary": "A new 16 GB card in a plain AM5 tower. Runs every 14B model at Q4 with long context, and gpt-oss-20b at full speed.",
   "reference_models": [
    "qwen3-14b",
    "gpt-oss-20b",
    "llama-3.1-8b"
   ],
   "idle_watts": 45,
   "parts": [
    {
     "role": "GPU",
     "name": "GeForce RTX 5060 Ti 16 GB",
     "usd": 479
    },
    {
     "role": "CPU",
     "name": "Ryzen 5 7600",
     "usd": 189
    },
    {
     "role": "Motherboard",
     "name": "B650 micro-ATX",
     "usd": 129
    },
    {
     "role": "Memory",
     "name": "32 GB DDR5-6000 (2×16)",
     "usd": 109
    },
    {
     "role": "Storage",
     "name": "2 TB NVMe Gen4",
     "usd": 119
    },
    {
     "role": "PSU",
     "name": "650 W 80+ Gold",
     "usd": 79
    },
    {
     "role": "Case",
     "name": "Mesh mid-tower",
     "usd": 69
    }
   ],
   "flag": "sample",
   "note": "Sample parts and prices for layout; not tracked retailer prices."
  },
  {
   "id": "used-3090-24gb",
   "name": "Used-3090 24 GB",
   "label": "Editors' pick",
   "hardware": "rtx-3090-24g-used",
   "summary": "The reference rig. A used RTX 3090 gives 24 GB at 936 GB/s, enough for every 24–32B model at Q4.",
   "reference_models": [
    "gemma-3-27b",
    "qwen3-30b-a3b",
    "mistral-small-3.2-24b"
   ],
   "idle_watts": 70,
   "parts": [
    {
     "role": "GPU",
     "name": "GeForce RTX 3090 24 GB (used",
     "eBay sold median)": null,
     "usd": 1200,
     "flag": "sourced",
     "source": "https://www.getpcparts.com/market-prices/gpu-graphics-cards/rtx-3090"
    },
    {
     "role": "CPU",
     "name": "Ryzen 5 7600",
     "usd": 189
    },
    {
     "role": "Motherboard",
     "name": "B650 ATX",
     "usd": 149
    },
    {
     "role": "Memory",
     "name": "64 GB DDR5-6000 (2×32)",
     "usd": 189
    },
    {
     "role": "Storage",
     "name": "2 TB NVMe Gen4",
     "usd": 119
    },
    {
     "role": "PSU",
     "name": "850 W 80+ Gold",
     "usd": 109
    },
    {
     "role": "Case",
     "name": "Mesh mid-tower",
     "usd": 25
    }
   ],
   "flag": "sample",
   "note": "GPU price is sourced (eBay sold, Oct 2026); other parts are sample."
  },
  {
   "id": "mac-studio-unified",
   "name": "Mac Studio M5 Max",
   "label": "Quietest",
   "hardware": "mac-studio-m5-max",
   "summary": "No build, no noise. The base Mac Studio runs 27B-class models at Q4 and sits silently on a desk.",
   "reference_models": [
    "gemma-3-27b",
    "qwen3-14b",
    "gpt-oss-20b"
   ],
   "idle_watts": 12,
   "parts": [
    {
     "role": "System",
     "name": "Mac Studio M5 Max (base configuration)",
     "usd": 2499,
     "flag": "sourced",
     "source": "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/"
    }
   ],
   "flag": "sample",
   "note": "Price is sourced (Apple, Aug 2026); base memory and bandwidth are sample."
  },
  {
   "id": "strix-halo-128gb",
   "name": "Strix Halo SFF 128 GB",
   "label": "Most memory per dollar",
   "hardware": "ryzen-ai-max-395-128g",
   "summary": "128 GB of unified memory in a mini PC. Holds gpt-oss-120b with room to spare; dense 70B models are slow.",
   "reference_models": [
    "gpt-oss-120b",
    "qwen3-30b-a3b",
    "llama-3.3-70b"
   ],
   "idle_watts": 15,
   "parts": [
    {
     "role": "System",
     "name": "Ryzen AI Max+ 395 mini PC",
     "128 GB": null,
     "usd": 3499.99,
     "flag": "sourced",
     "source": "https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/"
    }
   ],
   "flag": "sourced",
   "note": "Price from CFG (Oct 2026); 120 W load is the brief's break-even assumption."
  },
  {
   "id": "dual-gpu-48gb",
   "name": "Dual-GPU 48 GB",
   "label": "70B at home",
   "hardware": "2x-rtx-3090-48g-used",
   "summary": "Two used 3090s split a 70B model by layers. Capacity doubles; single-stream bandwidth does not.",
   "reference_models": [
    "llama-3.3-70b",
    "qwen3-32b",
    "gemma-3-27b"
   ],
   "idle_watts": 110,
   "parts": [
    {
     "role": "GPU",
     "name": "2× GeForce RTX 3090 24 GB (used)",
     "usd": 2400
    },
    {
     "role": "CPU",
     "name": "Ryzen 9 7900",
     "usd": 349
    },
    {
     "role": "Motherboard",
     "name": "X670E ATX (x8/x8)",
     "usd": 299
    },
    {
     "role": "Memory",
     "name": "128 GB DDR5-5600 (4×32)",
     "usd": 359
    },
    {
     "role": "Storage",
     "name": "4 TB NVMe Gen4",
     "usd": 229
    },
    {
     "role": "PSU",
     "name": 1,
     "300 W 80+ Platinum": null,
     "usd": 249
    },
    {
     "role": "Case",
     "name": "Full tower",
     "airflow": null,
     "usd": 159
    }
   ],
   "flag": "sample",
   "note": "Sample parts and prices for layout."
  },
  {
   "id": "silent-home-server",
   "name": "Silent home server",
   "label": "Always on",
   "hardware": "rtx-4060-ti-16g",
   "summary": "A low-power 16 GB box for an always-on household assistant. Tuned for idle draw, not peak speed.",
   "reference_models": [
    "qwen3-8b",
    "gpt-oss-20b",
    "qwen3-14b"
   ],
   "idle_watts": 30,
   "parts": [
    {
     "role": "GPU",
     "name": "GeForce RTX 4060 Ti 16 GB (power-limited to 150 W)",
     "usd": 449
    },
    {
     "role": "CPU",
     "name": "Ryzen 5 8600G",
     "usd": 179
    },
    {
     "role": "Motherboard",
     "name": "B650 mini-ITX",
     "usd": 169
    },
    {
     "role": "Memory",
     "name": "32 GB DDR5-5600 (2×16)",
     "usd": 99
    },
    {
     "role": "Storage",
     "name": "2 TB NVMe Gen4",
     "usd": 119
    },
    {
     "role": "PSU",
     "name": "550 W SFX 80+ Platinum",
     "usd": 129
    },
    {
     "role": "Case",
     "name": "Fanless-friendly ITX case",
     "usd": 139
    }
   ],
   "flag": "sample",
   "note": "Sample parts and prices for layout."
  }
 ],
 "runs": [
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-5090-32g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 58.2,
   "min": 56.9,
   "max": 59,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-4090-24g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 33.4,
   "min": 32.8,
   "max": 33.9,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 30.2,
   "min": 29.8,
   "max": 30.6,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rx-7900-xtx-24g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (Vulkan)",
   "test": "tg128 @ 8k depth",
   "tok_s": 27.1,
   "min": 26.2,
   "max": 27.6,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (Metal)",
   "test": "tg128 @ 8k depth",
   "tok_s": 13.9,
   "min": 13.5,
   "max": 14.2,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "dgx-spark-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 7.4,
   "min": 7.2,
   "max": 7.6,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (Vulkan)",
   "test": "tg128 @ 8k depth",
   "tok_s": 6.9,
   "min": 6.6,
   "max": 7.1,
   "repeats": 5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 0",
   "depth": 0,
   "tok_s": 30.2,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 4096",
   "depth": 4096,
   "tok_s": 29.1,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 16384",
   "depth": 16384,
   "tok_s": 26,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 32768",
   "depth": 32768,
   "tok_s": 23.1,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 0",
   "depth": 0,
   "tok_s": 13.9,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 4096",
   "depth": 4096,
   "tok_s": 13.1,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 16384",
   "depth": 16384,
   "tok_s": 11.2,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 32768",
   "depth": 32768,
   "tok_s": 9.4,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 0",
   "depth": 0,
   "tok_s": 6.9,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 4096",
   "depth": 4096,
   "tok_s": 6.5,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 16384",
   "depth": 16384,
   "tok_s": 5.6,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gemma-3-27b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp",
   "test": "tg128 @ depth 32768",
   "depth": 32768,
   "tok_s": 4.8,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder depth curve from the mockup."
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "rtx-3090-24g-used",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 128.4,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "qwen3-30b-a3b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 112,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "qwen3-14b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 52.3,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "mistral-small-3.2-24b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_K_M",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128 @ 8k depth",
   "tok_s": 34.1,
   "flag": "sample",
   "date": "2026-10-12",
   "note": "Placeholder figure from the approved mockup; not a measurement."
  },
  {
   "model": "llama-2-7b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_0",
   "runtime": "llamafile (Metal)",
   "test": "LocalScore generation",
   "tok_s": 62.5,
   "flag": "sourced",
   "source": "https://www.localscore.ai/model/1248",
   "date": "2026-10-12",
   "note": "Third-party measurement used to calibrate η (brief §4.3): implies η ≈ 0.43."
  },
  {
   "model": "llama-2-7b",
   "hardware": "rtx-5090-32g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 290.02,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 288.92,
   "max": 291.12
  },
  {
   "model": "llama-2-7b",
   "hardware": "rtx-4090-24g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 186.21,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 186.08,
   "max": 186.34
  },
  {
   "model": "llama-2-7b",
   "hardware": "rtx-3090-24g-used",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 158.16,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 157.95,
   "max": 158.37
  },
  {
   "model": "llama-2-7b",
   "hardware": "rtx-5060-ti-16g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 90.94,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 90.92,
   "max": 90.96
  },
  {
   "model": "llama-2-7b",
   "hardware": "rtx-4060-ti-16g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 63.86,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Measured on the 8 GB variant (same 288 GB/s bus); treat as indicative for the 16 GB card.",
   "min": 63.85,
   "max": 63.87
  },
  {
   "model": "llama-2-7b",
   "hardware": "dgx-spark-128g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg128",
   "tok_s": 57.21,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 57.15,
   "max": 57.27
  },
  {
   "model": "llama-2-7b",
   "hardware": "rx-7900-xtx-24g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (Vulkan)",
   "test": "tg128",
   "tok_s": 182.63,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.",
   "min": 182.41,
   "max": 182.85
  },
  {
   "model": "llama-2-7b",
   "hardware": "ryzen-ai-max-395-128g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (Vulkan)",
   "test": "tg128",
   "tok_s": 51.81,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Device reported as \"AMD Ryzen AI Max+ 300 Series\" (Radeon 8060S).",
   "min": 51.68,
   "max": 51.94
  },
  {
   "model": "llama-2-7b",
   "hardware": "macbook-pro-m4-max-128g",
   "quant": "Q4_0",
   "runtime": "llama.cpp (Metal)",
   "test": "tg128",
   "tok_s": 83.06,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, M4 Max 40-core GPU (546 GB/s), pinned commit 8e672ef (Nov 2023) for comparability; newer builds are faster."
  },
  {
   "model": "llama-2-7b",
   "hardware": "mac-studio-m5-max",
   "quant": "Q4_0",
   "runtime": "llama.cpp (Metal)",
   "test": "tg128",
   "tok_s": 119.92,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, measured on the M5 Max 40-core GPU (614 GB/s) option, commit c1d0e7a. The base 32-core (460 GB/s) model will be slower."
  },
  {
   "model": "llama-2-7b",
   "hardware": "mac-studio-m5-ultra",
   "quant": "Q4_0",
   "runtime": "llama.cpp (Metal)",
   "test": "tg128",
   "tok_s": 179.1,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
   "note": "Llama 2 7B Q4_0, llama-bench tg128, measured on the M5 Ultra 80-core GPU option (1.2 TB/s), commit c1d0e7a. The base 64-core model may be slightly slower."
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 0",
   "tok_s": 83.43,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 82.84,
   "max": 84.02,
   "depth": 0
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 4096",
   "tok_s": 79.22,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 78.62,
   "max": 79.82,
   "depth": 4096
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 8192",
   "tok_s": 75.22,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 74.17,
   "max": 76.27,
   "depth": 8192
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 16384",
   "tok_s": 70.36,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 69.99,
   "max": 70.73,
   "depth": 16384
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 32768",
   "tok_s": 61.65,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 61.35,
   "max": 61.95,
   "depth": 32768
  },
  {
   "model": "gpt-oss-20b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ 8k depth",
   "tok_s": 75.22,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 74.17,
   "max": 76.27
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 0",
   "tok_s": 58.72,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 58.52,
   "max": 58.92,
   "depth": 0
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 4096",
   "tok_s": 55.67,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 55.32,
   "max": 56.02,
   "depth": 4096
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 8192",
   "tok_s": 52.87,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 52.44,
   "max": 53.3,
   "depth": 8192
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 16384",
   "tok_s": 49.45,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 49.25,
   "max": 49.65,
   "depth": 16384
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ depth 32768",
   "tok_s": 42.76,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 42.62,
   "max": 42.9,
   "depth": 32768
  },
  {
   "model": "gpt-oss-120b",
   "hardware": "dgx-spark-128g",
   "quant": "MXFP4",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ 8k depth",
   "tok_s": 52.87,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.",
   "min": 52.44,
   "max": 53.3
  },
  {
   "model": "qwen3-coder-30b-a3b-instruct",
   "hardware": "dgx-spark-128g",
   "quant": "Q8_0",
   "runtime": "llama.cpp (CUDA)",
   "test": "tg32 @ 8k depth",
   "tok_s": 48.02,
   "flag": "sourced",
   "date": "2026-10-12",
   "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
   "note": "llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Model is Qwen3-Coder-30B-A3B-Instruct. Catalogue correction 2026-10-13: model identifier corrected from Qwen3-30B-A3B; shared architecture does not establish identical performance.",
   "min": 47.62,
   "max": 48.42
  }
 ],
 "plans": {
  "as_of": "2026-10-12",
  "flag": "sourced",
  "note": "US monthly list prices from each vendor's own pricing page, checked 2026-10-12. CAD shown where the vendor page showed it; otherwise USD × 1.4271 (Bank of Canada, 2026-10-09), flagged estimate.",
  "plans": [
   {
    "id": "chatgpt-plus",
    "tier": "standard",
    "name": "ChatGPT Plus",
    "vendor": "OpenAI",
    "usd": 20,
    "source": "https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus",
    "flag": "sourced",
    "cad": 28.54,
    "cad_flag": "estimate"
   },
   {
    "id": "claude-pro",
    "tier": "standard",
    "name": "Claude Pro",
    "vendor": "Anthropic",
    "usd": 20,
    "source": "https://claude.com/pricing",
    "note": "$20 billed monthly ($17/mo billed annually).",
    "flag": "sourced",
    "cad": 28.54,
    "cad_flag": "estimate"
   },
   {
    "id": "google-ai-pro",
    "tier": "standard",
    "name": "Google AI Pro",
    "vendor": "Google",
    "usd": 19.99,
    "cad": 26.99,
    "source": "https://gemini.google/subscriptions/",
    "cad_source": "https://one.google.com/about/plans",
    "flag": "sourced",
    "cad_flag": "sourced"
   },
   {
    "id": "claude-max-5x",
    "tier": "heavy",
    "name": "Claude Max (5×)",
    "vendor": "Anthropic",
    "usd": 100,
    "source": "https://claude.com/pricing",
    "flag": "sourced",
    "cad": 142.71,
    "cad_flag": "estimate"
   },
   {
    "id": "chatgpt-pro-100",
    "tier": "heavy",
    "name": "ChatGPT Pro 100",
    "vendor": "OpenAI",
    "usd": 100,
    "source": "https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers",
    "flag": "sourced",
    "cad": 142.71,
    "cad_flag": "estimate"
   },
   {
    "id": "google-ai-ultra",
    "tier": "heavy",
    "name": "Google AI Ultra (5×)",
    "vendor": "Google",
    "usd": 99.99,
    "source": "https://gemini.google/subscriptions/",
    "flag": "sourced",
    "cad": 142.7,
    "cad_flag": "estimate"
   },
   {
    "id": "claude-max-20x",
    "tier": "heavy",
    "name": "Claude Max (20×)",
    "vendor": "Anthropic",
    "usd": 200,
    "source": "https://claude.com/pricing",
    "flag": "sourced",
    "cad": 285.42,
    "cad_flag": "estimate"
   },
   {
    "id": "chatgpt-pro-200",
    "tier": "heavy",
    "name": "ChatGPT Pro 200",
    "vendor": "OpenAI",
    "usd": 200,
    "source": "https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers",
    "flag": "sourced",
    "cad": 285.42,
    "cad_flag": "estimate"
   },
   {
    "id": "google-ai-ultra-top",
    "tier": "heavy",
    "name": "Google AI Ultra (20×)",
    "vendor": "Google",
    "usd": 199.99,
    "source": "https://gemini.google/subscriptions/",
    "flag": "sourced",
    "cad": 285.41,
    "cad_flag": "estimate"
   },
   {
    "id": "chatgpt-pro",
    "tier": "heavy",
    "name": "ChatGPT Pro 500 (top tier)",
    "vendor": "OpenAI",
    "usd": 500,
    "source": "https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers",
    "flag": "sourced",
    "cad": 713.55,
    "cad_flag": "estimate"
   }
  ]
 },
 "glossary": [
  {
   "term": "AWQ and GPTQ",
   "slug": "awq-gptq",
   "definition": "Calibrated, mostly 4-bit GPU quantization methods. GPTQ uses approximate second-order information layer by layer; AWQ protects the roughly 1% of weights that matter most to activations. Both are first-class in vLLM and SGLang.",
   "seen_as": [
    "-AWQ",
    "-GPTQ",
    "w4a16"
   ],
   "explainer": "quantization-formats",
   "related": [
    "quantization",
    "gguf"
   ]
  },
  {
   "term": "Bits per weight (bpw)",
   "slug": "bits-per-weight",
   "definition": "The average number of bits used to store each parameter in a quantized file, including scales and metadata. Q4_K_M averages about 4.85 bpw, not 4. Multiply parameters by bpw and divide by 8 to estimate the file size in bytes.",
   "seen_as": [
    "4.85 bpw",
    "BPW"
   ],
   "explainer": "quantization",
   "related": [
    "quantization",
    "k-quants",
    "gguf"
   ]
  },
  {
   "term": "Context window",
   "slug": "context-window",
   "definition": "The maximum number of tokens a model can attend to at once, prompt and reply combined. A longer window does not cost memory until you use it; the KV cache grows with the tokens actually in context.",
   "seen_as": [
    "128K",
    "--ctx-size 32768",
    "num_ctx"
   ],
   "explainer": "kv-cache",
   "related": [
    "kv-cache",
    "rope-scaling",
    "sliding-window-attention"
   ]
  },
  {
   "term": "Decode",
   "slug": "decode",
   "definition": "The phase where the model generates the reply one token at a time. Each token requires reading the active weights and the KV cache, so decode speed (tok/s) is mostly set by memory bandwidth. llama-bench reports it as tg128.",
   "seen_as": [
    "tg128",
    "eval rate",
    "tok/s"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "prefill",
    "memory-bandwidth",
    "ttft"
   ]
  },
  {
   "term": "Embeddings",
   "slug": "embeddings",
   "definition": "Lists of numbers that represent the meaning of a text, made by a small embedding model. Similar meanings give nearby vectors, which is how RAG finds relevant passages without matching keywords.",
   "seen_as": [
    "nomic-embed-text",
    "/api/embed",
    "--embedding"
   ],
   "explainer": "rag-and-embeddings",
   "related": [
    "rag"
   ]
  },
  {
   "term": "EXL2 / EXL3",
   "slug": "exl2",
   "definition": "Quantized formats for the ExLlama runtimes on NVIDIA GPUs. EXL2 mixes 2- to 8-bit pieces to hit any average bits per weight; EXL3, from ExLlamaV3, is based on QTIP. Usually served through TabbyAPI.",
   "seen_as": [
    "-exl2",
    "-exl3",
    "4.65bpw"
   ],
   "explainer": "quantization-formats",
   "related": [
    "quantization",
    "bits-per-weight"
   ]
  },
  {
   "term": "Flash attention",
   "slug": "flash-attention",
   "definition": "An attention algorithm that computes results in tiles without writing the full attention matrix to memory. In llama.cpp it reduces the compute buffer, speeds up long contexts and is required to quantize the V cache.",
   "seen_as": [
    "-fa",
    "--flash-attn on"
   ],
   "explainer": "kv-cache",
   "related": [
    "kv-cache"
   ]
  },
  {
   "term": "GGUF",
   "slug": "gguf",
   "definition": "The single-file model format used by llama.cpp and the tools built on it (Ollama, LM Studio, llamafile). It stores the quantized tensors plus metadata such as the architecture, tokenizer and quant type, so a runtime can load it without extra config files.",
   "seen_as": [
    "model-Q4_K_M.gguf",
    "-m model.gguf"
   ],
   "explainer": "quantization",
   "related": [
    "quantization",
    "k-quants",
    "llama-cpp"
   ]
  },
  {
   "term": "Grouped-query attention (GQA)",
   "slug": "gqa",
   "definition": "An attention design where several query heads share one key/value head. Llama 3.1 8B has 32 query heads but 8 KV heads, which cuts the KV cache to a quarter of what full multi-head attention would need.",
   "seen_as": [
    "num_key_value_heads",
    "n_head_kv"
   ],
   "explainer": "kv-cache",
   "related": [
    "kv-cache",
    "mla"
   ]
  },
  {
   "term": "Importance matrix (imatrix)",
   "slug": "imatrix",
   "definition": "A calibration file that records which weights matter most on sample text. Quantizing with an imatrix spends precision where it counts, which noticeably improves quality at 3 and 4 bits for the same file size.",
   "seen_as": [
    "--imatrix",
    "i1-Q4_K_M",
    "IQ4_XS"
   ],
   "explainer": "quantization",
   "related": [
    "k-quants",
    "quantization",
    "kl-divergence"
   ]
  },
  {
   "term": "K-quants",
   "slug": "k-quants",
   "definition": "llama.cpp's family of block quantization schemes (Q3_K, Q4_K, Q5_K, Q6_K). Weights are grouped into super-blocks with their own scales, and the _S/_M/_L suffix says how many sensitive tensors are kept at higher precision.",
   "seen_as": [
    "Q4_K_M",
    "Q5_K_S",
    "Q6_K"
   ],
   "explainer": "quantization",
   "related": [
    "quantization",
    "bits-per-weight",
    "imatrix"
   ]
  },
  {
   "term": "KL-divergence",
   "slug": "kl-divergence",
   "definition": "A measure of how far a quantized model's next-token probabilities drift from the full-precision model's on the same text. Zero means identical. It is a more direct test of quantization damage than perplexity, and is how The Local Bench reports quant fidelity.",
   "seen_as": [
    "--kl-divergence",
    "KLD 0.017"
   ],
   "explainer": "quantization",
   "related": [
    "quantization",
    "perplexity"
   ]
  },
  {
   "term": "KV cache",
   "slug": "kv-cache",
   "definition": "The memory where a transformer keeps the attention keys and values for every token it has already read, so it doesn't recompute them. It grows linearly with context length and with the number of concurrent users, and on long contexts it can rival the weights in size.",
   "seen_as": [
    "--ctx-size",
    "-ctk q8_0",
    "--cache-type-k"
   ],
   "explainer": "kv-cache",
   "related": [
    "context-window",
    "gqa",
    "sliding-window-attention",
    "flash-attention"
   ]
  },
  {
   "term": "MCP (Model Context Protocol)",
   "slug": "mcp",
   "definition": "An open protocol, using JSON-RPC 2.0, that lets AI apps connect to servers offering tools, resources and prompts. Any MCP-capable app can then use any MCP server, local or remote.",
   "seen_as": [
    "MCP server",
    "mcpServers",
    "stdio"
   ],
   "explainer": "tool-calling-and-mcp",
   "related": [
    "tool-calling"
   ]
  },
  {
   "term": "Memory bandwidth",
   "slug": "memory-bandwidth",
   "definition": "How many gigabytes per second the processor can read from its memory. For single-user text generation it is usually the speed limit, because every new token requires reading all active weights once. Ceiling tok/s ≈ bandwidth ÷ gigabytes read per token.",
   "seen_as": [
    "GB/s",
    "TB/s",
    "memory bus"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "decode",
    "unified-memory",
    "vram",
    "mixture-of-experts"
   ]
  },
  {
   "term": "Mixture of experts (MoE)",
   "slug": "mixture-of-experts",
   "definition": "A model split into many expert sub-networks, of which only a few run for each token. You must hold all of them in memory, but only the active ones are read per token, so a 30B MoE with 3B active decodes about as fast as a 3B dense model.",
   "seen_as": [
    "30B-A3B",
    "num_experts",
    "experts_per_token"
   ],
   "explainer": "mixture-of-experts",
   "related": [
    "memory-bandwidth",
    "decode"
   ]
  },
  {
   "term": "Multi-head latent attention (MLA)",
   "slug": "mla",
   "definition": "DeepSeek's attention design, which compresses keys and values into a small shared latent vector per token. It shrinks the KV cache far more than GQA, so memory per token depends on the latent rank rather than the number of heads.",
   "seen_as": [
    "kv_lora_rank",
    "qk_rope_head_dim"
   ],
   "explainer": "kv-cache",
   "related": [
    "gqa",
    "kv-cache"
   ]
  },
  {
   "term": "Offloading",
   "slug": "offloading",
   "definition": "Running some layers on the GPU and the rest on the CPU when a model doesn't fit in VRAM. It works, but each token's time is the sum of both parts, so even a small CPU share can halve the speed.",
   "seen_as": [
    "-ngl 40",
    "--n-gpu-layers",
    "num_gpu"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "vram",
    "memory-bandwidth"
   ]
  },
  {
   "term": "Perplexity",
   "slug": "perplexity",
   "definition": "How surprised a model is by a test text, on average; lower is better. Useful for comparing quants of the same model, but not across models with different tokenizers, and it can hide errors that KL-divergence catches.",
   "seen_as": [
    "PPL",
    "llama-perplexity"
   ],
   "explainer": null,
   "related": [
    "kl-divergence",
    "quantization"
   ]
  },
  {
   "term": "Prefill",
   "slug": "prefill",
   "definition": "The phase where the model reads your whole prompt in parallel before replying. It is limited by compute (TFLOPS) rather than bandwidth, and it sets how long you wait before the first token on long prompts. llama-bench reports it as pp512.",
   "seen_as": [
    "pp512",
    "prompt eval rate"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "decode",
    "ttft"
   ]
  },
  {
   "term": "Quantization",
   "slug": "quantization",
   "definition": "Storing a model's weights with fewer bits per number, for example 4 instead of 16. It shrinks the file and the memory needed roughly in proportion, and speeds up generation on bandwidth-limited hardware, at the cost of a small, measurable loss of accuracy.",
   "seen_as": [
    "Q4_K_M",
    "Q8_0",
    "MXFP4",
    "--quantize"
   ],
   "explainer": "quantization",
   "related": [
    "bits-per-weight",
    "gguf",
    "k-quants",
    "kl-divergence"
   ]
  },
  {
   "term": "RAG",
   "slug": "rag",
   "definition": "Retrieval-augmented generation. Your documents are split into chunks and embedded; the chunks closest to a question are pasted into the prompt so the model answers from them. Nothing is retrained.",
   "seen_as": [
    "knowledge base",
    "#documents",
    "top-k"
   ],
   "explainer": "rag-and-embeddings",
   "related": [
    "embeddings",
    "context-window"
   ]
  },
  {
   "term": "RoPE scaling (YaRN)",
   "slug": "rope-scaling",
   "definition": "Techniques that stretch a model's positional encoding to run beyond its trained context, such as Qwen3's 32K native window extended to 128K with YaRN. Quality can drop on short prompts if scaling is left on unnecessarily.",
   "seen_as": [
    "--rope-scaling yarn",
    "rope_scaling"
   ],
   "explainer": "context-length",
   "related": [
    "context-window"
   ]
  },
  {
   "term": "Sliding-window attention",
   "slug": "sliding-window-attention",
   "definition": "Layers that only attend to the most recent N tokens instead of the whole context. Their KV cache stops growing at N tokens. Gemma 3 uses a 1,024-token window on five of every six layers; gpt-oss uses 128 tokens on half of its layers.",
   "seen_as": [
    "sliding_window",
    "SWA",
    "--swa-full"
   ],
   "explainer": "kv-cache",
   "related": [
    "kv-cache",
    "context-window"
   ]
  },
  {
   "term": "Speculative decoding",
   "slug": "speculative-decoding",
   "definition": "A speed-up where a small draft model guesses several tokens and the large model verifies them in one pass. Accepted guesses come almost free, and the output is identical to the large model's. Typical gains are 1.5–3× on predictable text.",
   "seen_as": [
    "-md",
    "--spec-draft-n-max",
    "draft model",
    "EAGLE"
   ],
   "explainer": "speculative-decoding",
   "related": [
    "decode",
    "memory-bandwidth"
   ]
  },
  {
   "term": "Time to first token (TTFT)",
   "slug": "ttft",
   "definition": "The delay between sending a prompt and seeing the first token of the reply. Roughly prompt tokens ÷ prefill speed. It matters most for RAG and coding agents, which send prompts of 8k–16k tokens.",
   "seen_as": [
    "TTFT",
    "first token latency"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "prefill",
    "decode"
   ]
  },
  {
   "term": "Tool calling",
   "slug": "tool-calling",
   "definition": "A model's ability to answer with a structured request to run a function, such as a search or a calculator, instead of plain text. The app runs the function and feeds the result back. It needs a model and chat template trained for it.",
   "seen_as": [
    "function calling",
    "tools",
    "tool_calls",
    "--jinja"
   ],
   "explainer": "tool-calling-and-mcp",
   "related": [
    "mcp"
   ]
  },
  {
   "term": "Unified memory",
   "slug": "unified-memory",
   "definition": "A single pool of memory shared by CPU and GPU, as on Apple Silicon, AMD Strix Halo and NVIDIA DGX Spark. It offers large capacity (up to 512 GB) but the OS reserves part of it; we assume 75% is usable by the GPU by default.",
   "seen_as": [
    "128 GB unified",
    "iogpu.wired_limit_mb"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "vram",
    "memory-bandwidth"
   ]
  },
  {
   "term": "VRAM",
   "slug": "vram",
   "definition": "The memory on a discrete graphics card. A model runs at full speed only if its weights, KV cache and runtime overhead all fit in VRAM; whatever spills over to system RAM runs at system-RAM speed.",
   "seen_as": [
    "24 GB",
    "-ngl 99",
    "out of memory"
   ],
   "explainer": "memory-bandwidth",
   "related": [
    "unified-memory",
    "offloading",
    "memory-bandwidth"
   ]
  }
 ]
}