{
  "title": "How long you stare at the cursor",
  "footnote": "Medians of five runs at 350 W. Same prompts in both phases, different per run, so prompt caching can't inflate this.",
  "kind": "dumbbell",
  "options": {
    "min": 0,
    "max": 32
  },
  "rows": [
    {
      "label": "Qwen3.5-9B",
      "note": "BF16 on vLLM, Strix-B",
      "a": 8.448,
      "b": 5.415,
      "delta": "-3.0s",
      "_group": "Unquantized or vLLM"
    },
    {
      "label": "Qwen3-VL-8B",
      "note": "INT4 on vLLM, PNY",
      "a": 12.123,
      "b": 8.394,
      "delta": "-3.7s",
      "_group": "Unquantized or vLLM"
    },
    {
      "label": "Gemma 4 E4B",
      "note": "BF16 on vLLM, Strix-B",
      "a": 6.733,
      "b": 5.345,
      "delta": "-1.4s",
      "_group": "Unquantized or vLLM"
    },
    {
      "label": "Ornith-1.5-9B",
      "note": "BF16 GGUF, Zotac",
      "a": 12.109,
      "b": 8.998,
      "delta": "-3.1s",
      "_group": "Unquantized or vLLM"
    },
    {
      "label": "Gemma 4 12B",
      "note": "BF16 on vLLM, 2 cards",
      "a": 16.295,
      "b": 14.081,
      "delta": "-2.2s",
      "_group": "Split across cards"
    },
    {
      "label": "Qwen3.8-27B",
      "note": "BF16 on vLLM, 4 cards",
      "a": 30.237,
      "b": 27.12,
      "delta": "-3.1s",
      "_group": "Split across cards"
    },
    {
      "label": "Qwen3.8-27B",
      "note": "Q4 GGUF, Strix-B",
      "a": 27.907,
      "b": 27.339,
      "delta": "-0.6s",
      "_group": "Q4 GGUF on llama.cpp"
    },
    {
      "label": "Gemma 4 12B QAT",
      "note": "Q4 GGUF, PNY",
      "a": 13.907,
      "b": 13.297,
      "delta": "-0.6s",
      "_group": "Q4 GGUF on llama.cpp"
    },
    {
      "label": "Ornith-1.5-35B",
      "note": "Q4 GGUF, Strix-B",
      "a": 9.335,
      "b": 9.129,
      "delta": "-0.2s",
      "_group": "Q4 GGUF on llama.cpp"
    }
  ],
  "series": [
    {
      "key": "a",
      "label": "Stock",
      "color": "muted",
      "suffix": "s",
      "decimals": 1
    },
    {
      "key": "b",
      "label": "Unlocked",
      "color": 1,
      "suffix": "s",
      "decimals": 1
    }
  ],
  "subtitle": "Time to first token on a 32K prompt, which is about what a coding agent sends once it has read a few files.",
  "__hash__": "z4YoFXqHD3GqKnJWNxaqQm9RaI_y8GnIKWA2fGk6jqk"
}