{
  "$comment": [
    "TTFT measured on one machine. REPORTED, NOT REPRODUCED — unlike every",
    "other figure in data/, CI cannot re-derive this, and a different machine",
    "will get different numbers. That is the point: TTFT is a property of a",
    "deployment. See data/ttft-sources.json.",
    "",
    "Re-run: node tools/measure-ttft.mjs --model <model>"
  ],
  "measuredAt": "2026-09-03",
  "reproducedByCI": false,
  "configurations": [
    {
      "id": "rtx3090-qwen3-27b-q4",
      "model": "huihui-qwen3.8-27b-abliterated:UD-Q4_K_XL",
      "accelerator": "NVIDIA GeForce RTX 3090",
      "acceleratorMemory": "24576 MiB",
      "driver": "595.84",
      "runtime": "ollama 0.32.14",
      "parameters": "27.3B",
      "quantisation": "Q4_K_M",
      "contextLength": 262144,
      "method": {
        "note": "Wall-clock from request to first streamed token, which is what a person waits.",
        "promptSource": "corpus/samples/ (CC0), seeded shuffle, unique nonce per run to defeat the prefix cache",
        "runsPerLength": 5,
        "thinking": "disabled, so \"first token\" means the same thing across models",
        "discardRule": "runs reporting load_duration > 500 ms had the model reloaded underneath them and are excluded"
      },
      "byPromptLength": [
        {
          "promptTokens": 200,
          "warm": {
            "n": 5,
            "medianMs": 843.7,
            "minMs": 780.7,
            "maxMs": 911.8
          },
          "prefillMedianMs": 398.2,
          "discardedForReload": 0
        },
        {
          "promptTokens": 1386,
          "warm": {
            "n": 5,
            "medianMs": 1931.3,
            "minMs": 1908.1,
            "maxMs": 2008.9
          },
          "prefillMedianMs": 1463.1,
          "discardedForReload": 0
        },
        {
          "promptTokens": 5445,
          "warm": {
            "n": 5,
            "medianMs": 5616.9,
            "minMs": 5418.3,
            "maxMs": 5688.4
          },
          "prefillMedianMs": 5022.9,
          "discardedForReload": 0
        },
        {
          "promptTokens": 10812,
          "warm": {
            "n": 5,
            "medianMs": 10967.3,
            "minMs": 10770.5,
            "maxMs": 11282.6
          },
          "prefillMedianMs": 10226.3,
          "discardedForReload": 0
        }
      ],
      "coldStart": {
        "n": 3,
        "medianMs": 13237.8,
        "minMs": 13179.8,
        "maxMs": 13307.7,
        "note": "model not resident; includes reading weights from disk"
      }
    }
  ]
}