{
  "charts": [
    {
      "metric": "metrics.decode_tok_s_mean",
      "series": [
        {
          "label": "Runtime arm",
          "points": [
            {
              "artifact": "capacity-nospec-32.json",
              "artifact_sha256": "6d78129ff56b22a2079cafce0087b047c9e605a5f30cc2faa22a20548e3381ae",
              "value": 71.38765771219772,
              "x": "No speculation"
            },
            {
              "artifact": "capacity-mtp3-32.json",
              "artifact_sha256": "f595470be6889a77d8915b8c087fa02397d690afd009a4a53c9de6f9e90c899f",
              "value": 176.1983312887445,
              "x": "MTP3"
            }
          ]
        }
      ],
      "title": "Mean decode",
      "x_label": "matched 8K/C1 strict 32-word",
      "y_label": "tok/s"
    },
    {
      "metric": "metrics.ttft_p50_ms",
      "series": [
        {
          "label": "Runtime arm",
          "points": [
            {
              "artifact": "capacity-nospec-32.json",
              "artifact_sha256": "6d78129ff56b22a2079cafce0087b047c9e605a5f30cc2faa22a20548e3381ae",
              "value": 264.19609994627535,
              "x": "No speculation"
            },
            {
              "artifact": "capacity-mtp3-32.json",
              "artifact_sha256": "f595470be6889a77d8915b8c087fa02397d690afd009a4a53c9de6f9e90c899f",
              "value": 268.71289999689907,
              "x": "MTP3"
            }
          ]
        }
      ],
      "title": "Median TTFT",
      "x_label": "matched 8K/C1 strict 32-word",
      "y_label": "ms"
    },
    {
      "metric": "metrics.e2e_p50_ms",
      "series": [
        {
          "label": "Runtime arm",
          "points": [
            {
              "artifact": "capacity-nospec-32.json",
              "artifact_sha256": "6d78129ff56b22a2079cafce0087b047c9e605a5f30cc2faa22a20548e3381ae",
              "value": 894.6875000256114,
              "x": "No speculation"
            },
            {
              "artifact": "capacity-mtp3-32.json",
              "artifact_sha256": "f595470be6889a77d8915b8c087fa02397d690afd009a4a53c9de6f9e90c899f",
              "value": 524.419300025329,
              "x": "MTP3"
            }
          ]
        }
      ],
      "title": "Median E2E",
      "x_label": "matched 8K/C1 strict 32-word",
      "y_label": "ms"
    },
    {
      "metric": "metrics.decode_tok_s_mean",
      "series": [
        {
          "label": "GGUF incumbent (separate runtime/quant)",
          "points": [
            {
              "artifact": "capacity-incumbent-32.json",
              "artifact_sha256": "ab32dfa71e92bdb53097f7ee830e110b256fac3dcdc22308d1986b10b2ec3e50",
              "value": 107.65218702787973,
              "x": "8K/C1"
            }
          ]
        },
        {
          "label": "Huihui MTP3",
          "points": [
            {
              "artifact": "capacity-mtp3-32.json",
              "artifact_sha256": "f595470be6889a77d8915b8c087fa02397d690afd009a4a53c9de6f9e90c899f",
              "value": 176.1983312887445,
              "x": "8K/C1"
            }
          ]
        }
      ],
      "title": "Cross-profile mean decode",
      "x_label": "8K/C1 strict 32-word; not a speculation control",
      "y_label": "tok/s"
    },
    {
      "metric": "metrics.e2e_mean_ms",
      "series": [
        {
          "label": "GGUF incumbent (separate runtime/quant)",
          "points": [
            {
              "artifact": "capacity-incumbent-32.json",
              "artifact_sha256": "ab32dfa71e92bdb53097f7ee830e110b256fac3dcdc22308d1986b10b2ec3e50",
              "value": 2571.788158330795,
              "x": "8K/C1"
            }
          ]
        },
        {
          "label": "Huihui MTP3",
          "points": [
            {
              "artifact": "capacity-mtp3-32.json",
              "artifact_sha256": "f595470be6889a77d8915b8c087fa02397d690afd009a4a53c9de6f9e90c899f",
              "value": 523.9915000080751,
              "x": "8K/C1"
            }
          ]
        }
      ],
      "title": "Cross-profile mean E2E",
      "x_label": "8K/C1 strict 32-word; not a speculation control",
      "y_label": "ms"
    }
  ],
  "metric_semantics": {
    "boundary": "No p99 claim. 32K is not plotted; GGUF is not speculation-control evidence.",
    "population": "12 strict controlled-output requests per arm"
  },
  "schema": "anvil-serving.benchmark-graph-data/v1",
  "subtitle": "NInfer no-spec/MTP3 pair is isolated. GGUF is a separately labeled cross-profile comparison with different checkpoint, runtime, and quantization.",
  "title": "Huihui NInfer compatible-runtime 8K/C1 strict 32-word comparison"
}