{
  "schema": "anvil-serving.benchmark-graph-manifest/v1",
  "title": "GLM-5.3-Flash: historical Windows / WSL2 versus native Linux",
  "subtitle": "Same model/image, TP2, C1; n=3 per point; variable output and shared-prefix prompts; whole-stack comparison",
  "charts": [
    {
      "title": "Decode median",
      "metric": "metrics.decode_tok_s_p50",
      "x_label": "Nominal context; actual prompt sizes match",
      "y_label": "tokens / second",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "x": "4k",
              "artifact": "windows-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "windows-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "windows-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "windows-capacity-380k-r3.json"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "x": "4k",
              "artifact": "linux-legacy-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "linux-legacy-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "linux-legacy-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "linux-legacy-capacity-380k-r3.json"
            }
          ]
        }
      ]
    },
    {
      "title": "Effective prefill median",
      "metric": "metrics.effective_prefill_tok_s_p50",
      "x_label": "Nominal context; actual prompt sizes match",
      "y_label": "tokens / second",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "x": "4k",
              "artifact": "windows-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "windows-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "windows-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "windows-capacity-380k-r3.json"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "x": "4k",
              "artifact": "linux-legacy-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "linux-legacy-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "linux-legacy-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "linux-legacy-capacity-380k-r3.json"
            }
          ]
        }
      ]
    },
    {
      "title": "Time to first visible token median",
      "metric": "metrics.ttft_p50_ms",
      "x_label": "Nominal context; actual prompt sizes match",
      "y_label": "milliseconds",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "x": "4k",
              "artifact": "windows-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "windows-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "windows-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "windows-capacity-380k-r3.json"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "x": "4k",
              "artifact": "linux-legacy-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "linux-legacy-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "linux-legacy-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "linux-legacy-capacity-380k-r3.json"
            }
          ]
        }
      ]
    },
    {
      "title": "End-to-end median",
      "metric": "metrics.e2e_p50_ms",
      "x_label": "Nominal context; actual prompt sizes match",
      "y_label": "milliseconds",
      "series": [
        {
          "label": "Windows / WSL2",
          "points": [
            {
              "x": "4k",
              "artifact": "windows-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "windows-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "windows-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "windows-capacity-380k-r3.json"
            }
          ]
        },
        {
          "label": "Native Linux",
          "points": [
            {
              "x": "4k",
              "artifact": "linux-legacy-capacity-4k-r3.json"
            },
            {
              "x": "120k",
              "artifact": "linux-legacy-capacity-120k-r3.json"
            },
            {
              "x": "262k",
              "artifact": "linux-legacy-capacity-262k-r3.json"
            },
            {
              "x": "380k",
              "artifact": "linux-legacy-capacity-380k-r3.json"
            }
          ]
        }
      ]
    }
  ]
}