{
  "title": "The longer the prompt, the more it matters",
  "footnote": "Prefill at 512 tokens is short enough that timing noise is high, so trust the 8K and 32K columns more.",
  "kind": "matrix",
  "options": {
    "columns": [
      {
        "key": "s",
        "label": "512"
      },
      {
        "key": "m",
        "label": "8K"
      },
      {
        "key": "l",
        "label": "32K"
      }
    ],
    "max": 65,
    "min": 0,
    "flatBelow": 3
  },
  "rows": [
    {
      "_group": "Prefill",
      "label": "Qwen3.5-9B",
      "note": "BF16, vLLM",
      "s": 43,
      "m": 54,
      "l": 56
    },
    {
      "_group": "Prefill",
      "label": "Gemma 4 E4B",
      "note": "BF16, vLLM",
      "s": 35.3,
      "m": 38.1,
      "l": 26
    },
    {
      "_group": "Prefill",
      "label": "Qwen3-VL-8B",
      "note": "INT4, vLLM",
      "s": 46.4,
      "m": 52.3,
      "l": 44.4
    },
    {
      "_group": "Prefill",
      "label": "Ornith-1.5-9B",
      "note": "BF16 GGUF",
      "s": 19.5,
      "m": 34.8,
      "l": 34.6
    },
    {
      "_group": "Prefill",
      "label": "Qwen3.8-27B",
      "note": "Q4 GGUF",
      "s": 0.3,
      "m": 0.3,
      "l": 2.1
    },
    {
      "_group": "Decode",
      "label": "Qwen3.5-9B",
      "note": "BF16, vLLM",
      "s": 3.6,
      "m": 3.9,
      "l": 4.4
    },
    {
      "_group": "Decode",
      "label": "Gemma 4 E4B",
      "note": "BF16, vLLM",
      "s": 4.8,
      "m": 4.5,
      "l": 5
    },
    {
      "_group": "Decode",
      "label": "Qwen3-VL-8B",
      "note": "INT4, vLLM",
      "s": 1.9,
      "m": 5.4,
      "l": 7.9
    },
    {
      "_group": "Decode",
      "label": "Ornith-1.5-9B",
      "note": "BF16 GGUF",
      "s": 0,
      "m": 0,
      "l": 0
    },
    {
      "_group": "Decode",
      "label": "Qwen3.8-27B",
      "note": "Q4 GGUF",
      "s": -0.1,
      "m": -0.3,
      "l": -0.6
    }
  ],
  "series": [
    {
      "key": "s",
      "label": "512",
      "suffix": "%",
      "decimals": 1
    },
    {
      "key": "m",
      "label": "8K",
      "suffix": "%",
      "decimals": 1
    },
    {
      "key": "l",
      "label": "32K",
      "suffix": "%",
      "decimals": 1
    }
  ],
  "subtitle": "Percent gain by prompt size in tokens, one card per model. Decode barely shows up until the context gets long.",
  "xLabel": "Model",
  "__hash__": "xtUD__l3TDcSeOaE9RXXnXRgVOhSZrg6UFqNJmJA_S8"
}