{
  "id": "inbox-latency",
  "story": "llm-bench-2026-09-27",
  "title": "The cost of the same inbox workload",
  "unit": "median seconds / task",
  "rows": [
    {
      "label": "A Swift-1.5 27B IQ3_S (Win)",
      "value": 51.2,
      "highlight": false
    },
    {
      "label": "B Bonsai-2-27B-Abl PQ2_0 (RTX 3060)",
      "value": 26.6,
      "highlight": false
    },
    {
      "label": "C Qwen3.8-27B GSQ-RCO IQ3_S (Win)",
      "value": 59.9,
      "highlight": false
    },
    {
      "label": "D Gemma4-12B Q4_K_M+MTP (RTX 3060)",
      "value": 8.0,
      "highlight": false
    },
    {
      "label": "E Xing4.0-29B-A4B IQ3_XXS (RTX 3060)",
      "value": 69.0,
      "highlight": false
    },
    {
      "label": "F MiMo-9B MLX 4bit (Mac)",
      "value": 6.7,
      "highlight": false
    },
    {
      "label": "F MiMo-9B MLX 4bit (Mac)",
      "value": 17.4,
      "highlight": false
    },
    {
      "label": "G Hermes-4-14B-OBL EXL3 4.0 (RTX 3060)",
      "value": 94.4,
      "highlight": false
    },
    {
      "label": "H Bonsai-2-27B MLX 2bit (Mac)",
      "value": 87.0,
      "highlight": false
    },
    {
      "label": "I Bonsai-2-27B-Abl PQ2_0 (Mac)",
      "value": 81.7,
      "highlight": false
    },
    {
      "label": "J54 Bonsai-2-27B PQ2_0+MTP (RTX 3060)",
      "value": 15.8,
      "highlight": true
    },
    {
      "label": "JMac Bonsai-2-27B PQ2_0+MTP (Mac)",
      "value": 67.1,
      "highlight": false
    },
    {
      "label": "L ThinkingCap-Qwen3.8-27B IQ3_S (Mac)",
      "value": 81.9,
      "highlight": false
    },
    {
      "label": "M ZDTaichu5.0-9B MLX 8bit (Mac, text-only)",
      "value": 65.0,
      "highlight": false
    },
    {
      "label": "K54 Bonsai-2 PQ2_0+DFlash2 (3060, CPU draft)",
      "value": 52.0,
      "highlight": false
    },
    {
      "label": "KMac Bonsai-2 PQ2_0+DFlash2 (Mac)",
      "value": 71.6,
      "highlight": false
    },
    {
      "label": "NMac HauhauCS Qwen3.8-27B IQ3_XS+MTP (Mac)",
      "value": 135.5,
      "highlight": false
    }
  ],
  "sources": [
    {
      "path": "llm-bench-2026-09-27/README.md",
      "url": "https://github.com/groxaxo/experimentos/blob/6550ead3945b7ecacbaac1e0767edc49a9de9747/llm-bench-2026-09-27/README.md",
      "sha256": "c8f1976fe45d4d104e9bc01ce0fd58923213a83ab202c0df57dc005848ee6157"
    }
  ],
  "revision": "6550ead3945b7ecacbaac1e0767edc49a9de9747",
  "scope": "Thinking-on and forced-thinking rows · same 50 tasks",
  "caveat": "Median client wall time. A higher pass rate does not automatically give a faster workflow.",
  "kind": "bar",
  "maximum": null,
  "threshold": null,
  "decimals": 1,
  "origin": "reported aggregate table"
}