{
  "id": "sharp-completion",
  "story": "qwen38-templates-2026-08-27-medium-xhigh-template-ab",
  "title": "Sharp reduced completion tokens in all three deployments",
  "unit": "completion-token reduction (%)",
  "rows": [
    {
      "label": "Hauhau Q6",
      "value": 20.6,
      "highlight": true
    },
    {
      "label": "Unsloth Q6",
      "value": 20.1,
      "highlight": true
    },
    {
      "label": "OrcaRouter FP8",
      "value": 16.0,
      "highlight": true
    }
  ],
  "sources": [
    {
      "path": "qwen38/templates/2026-08-27-medium-xhigh-template-ab/BLOG.md",
      "url": "https://github.com/groxaxo/experimentos/blob/6550ead3945b7ecacbaac1e0767edc49a9de9747/qwen38/templates/2026-08-27-medium-xhigh-template-ab/BLOG.md",
      "sha256": "b77a911508d7aaa39d26be33b4f6a48aefdec8249e132fe58ecf647d78a398e3"
    }
  ],
  "revision": "6550ead3945b7ecacbaac1e0767edc49a9de9747",
  "scope": "Medium + xhigh combined · stock vs Sharp v22.4.0",
  "caveat": "The longer prompt reduces the net saving to 9.0%, 6.5% and 3.2%. Small accuracy differences were not significant.",
  "kind": "bar",
  "maximum": null,
  "threshold": null,
  "decimals": 1,
  "origin": "reported paired aggregate table"
}