{
  "charts": [
    {
      "metric": "metrics.decode_tok_s_p50",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "artifact": "windows-endurance-4k-r60.json",
              "artifact_sha256": "655fbecec5d87e187b7253c9ae05ab64508270110a8bda8ded2f2a68d5a145cb",
              "value": 102.19275206464441,
              "x": "60 requests"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "artifact": "endurance-4k-r60.json",
              "artifact_sha256": "9effa5ad700996bf0c82264edc1a555b3791258170b8e4490729287f20c45ca6",
              "value": 142.6429205479183,
              "x": "60 requests"
            }
          ]
        }
      ],
      "title": "Decode median: higher is faster",
      "x_label": "4K nominal prompt target",
      "y_label": "tokens / second"
    },
    {
      "metric": "metrics.ttft_p50_ms",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "artifact": "windows-endurance-4k-r60.json",
              "artifact_sha256": "655fbecec5d87e187b7253c9ae05ab64508270110a8bda8ded2f2a68d5a145cb",
              "value": 183.42089996440336,
              "x": "60 requests"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "artifact": "endurance-4k-r60.json",
              "artifact_sha256": "9effa5ad700996bf0c82264edc1a555b3791258170b8e4490729287f20c45ca6",
              "value": 123.24624299799325,
              "x": "60 requests"
            }
          ]
        }
      ],
      "title": "First visible token median: lower is faster",
      "x_label": "4K nominal prompt target",
      "y_label": "milliseconds"
    }
  ],
  "metric_semantics": {},
  "schema": "anvil-serving.benchmark-graph-data/v1",
  "subtitle": "Warm direct, TP2/C1, 4K target; n=60 per platform; variable short outputs and shared prefixes",
  "title": "GLM-5.3-Flash: 60-request endurance comparison"
}