{
  "id": "context-decode-sweep",
  "story": "context-configured-versus-tested",
  "title": "Two draft tokens led this fixed-output sweep",
  "unit": "tokens / second",
  "rows": [
    {
      "label": "GPTQ INT4 · drafting off",
      "value": 45.3,
      "highlight": false
    },
    {
      "label": "GPTQ INT4 · MTP2",
      "value": 65.7,
      "highlight": true
    },
    {
      "label": "GPTQ INT4 · MTP3",
      "value": 61.6,
      "highlight": false
    },
    {
      "label": "CyanKiwi · base",
      "value": 38.9,
      "highlight": false
    }
  ],
  "scope": "One GPU · three fixed 512-output-token requests · configured context 8192",
  "caveat": "Decode-only sweep. Coding-task outcomes and context capacity require separate evidence.",
  "kind": "bar",
  "decimals": 2,
  "revision": "7e6fdbe3b1d1aa360346255bfc62945dba694399",
  "sources": [
    {
      "path": "qwen38/context-speed-audit-2026-10-06/README.md",
      "url": "https://github.com/groxaxo/experimentos/blob/7e6fdbe3b1d1aa360346255bfc62945dba694399/qwen38/context-speed-audit-2026-10-06/README.md",
      "sha256": "4cef8c9776e8409d50c57de25f41f96e28d846f8554b38f30acdb30db36022a1"
    }
  ],
  "origin": "reviewed report table",
  "maximum": null,
  "threshold": null
}