{
  "chart": "s2s-chart/1",
  "kind": "bars",
  "title": "Is a mixture of experts (MoE) really faster from RAM? A mixture beside a dense model on the same machine",
  "subtitle": "Tokens a second while writing, a mixture of experts beside a dense model on the same machine, at default threads and (third panel) at 20; a longer bar writes faster. Every gemma4:26b bar includes its built-in draft model; no dense model has one.",
  "unit": "tok/s",
  "series": [
    "The mixture",
    "The dense model"
  ],
  "panels": [
    {
      "heading": "26B mixture vs 24B dense, default threads",
      "rows": [
        {
          "label": "EPYC 7532, 128 GB",
          "notes": "Mixture over dense: 3.48",
          "values": [
            {
              "value": "24.68",
              "source": "/two-hours-on-battery/data/runs/cpu-server-m5.json derived.decode_tok_s.median 24.675"
            },
            {
              "value": "7.09",
              "source": "/two-hours-on-battery/data/runs/cpu-server-m4.json derived.decode_tok_s.median 7.091"
            }
          ]
        },
        {
          "label": "290HX Plus, 64 GB",
          "notes": "Mixture over dense: 4.91",
          "values": [
            {
              "value": "13.26",
              "source": "/two-hours-on-battery/data/runs/cpu-laptop-m5.json derived.decode_tok_s.median 13.258"
            },
            {
              "value": "2.70",
              "source": "/two-hours-on-battery/data/runs/cpu-laptop-m4.json derived.decode_tok_s.median 2.702"
            }
          ]
        },
        {
          "label": "i5-12600H, 32 GB",
          "notes": "Mixture over dense: 3.82",
          "values": [
            {
              "value": "12.66",
              "source": "/two-hours-on-battery/data/runs/cpu-mini-m5.json derived.decode_tok_s.median 12.657"
            },
            {
              "value": "3.31",
              "source": "/two-hours-on-battery/data/runs/cpu-mini-m4.json derived.decode_tok_s.median 3.313"
            }
          ]
        },
        {
          "label": "3700X, 32 GB DDR4",
          "notes": "Mixture over dense: 5.31",
          "values": [
            {
              "value": "13.64",
              "source": "/speeds-from-ram-alone/data/two-channel-desktop/summary.json arms[0].decode_median 13.642"
            },
            {
              "value": "2.57",
              "source": "/speeds-from-ram-alone/data/two-channel-desktop/summary.json arms[1].decode_median 2.569"
            }
          ]
        }
      ]
    },
    {
      "heading": "30B-A3B mixture vs 9B dense, default threads",
      "rows": [
        {
          "label": "EPYC 7532, 128 GB",
          "notes": "Mixture over dense: 2.31",
          "values": [
            {
              "value": "31.47",
              "source": "/two-hours-on-battery/data/runs/cpu-server-m6.json derived.decode_tok_s.median 31.469"
            },
            {
              "value": "13.59",
              "source": "/two-hours-on-battery/data/runs/cpu-server-m1.json derived.decode_tok_s.median 13.594"
            }
          ]
        },
        {
          "label": "290HX Plus, 64 GB",
          "notes": "Mixture over dense: 2.06",
          "values": [
            {
              "value": "15.35",
              "source": "/two-hours-on-battery/data/runs/cpu-laptop-m6.json derived.decode_tok_s.median 15.35"
            },
            {
              "value": "7.45",
              "source": "/two-hours-on-battery/data/runs/cpu-laptop-m1.json derived.decode_tok_s.median 7.45"
            }
          ]
        },
        {
          "label": "i5-12600H, 32 GB",
          "notes": "Mixture over dense: 1.60",
          "values": [
            {
              "value": "14.61",
              "source": "/two-hours-on-battery/data/runs/cpu-mini-m6.json derived.decode_tok_s.median 14.605"
            },
            {
              "value": "9.11",
              "source": "/two-hours-on-battery/data/runs/cpu-mini-m1.json derived.decode_tok_s.median 9.105"
            }
          ]
        }
      ]
    },
    {
      "heading": "The laptop at 20 threads (2026-08-25)",
      "rows": [
        {
          "label": "26B vs 24B dense",
          "notes": "Mixture over dense: 1.64",
          "values": [
            {
              "value": "8.44",
              "source": "/two-hours-on-battery/data/supplement-runs/tuned-cpu-laptop-m5.json derived.decode_tok_s.median 8.438 (num_thread 20)"
            },
            {
              "value": "5.14",
              "source": "/two-hours-on-battery/data/supplement-runs/tuned-cpu-laptop-m4.json derived.decode_tok_s.median 5.14 (num_thread 20)"
            }
          ]
        },
        {
          "label": "30B-A3B vs 9B",
          "notes": "Mixture over dense: 0.65",
          "values": [
            {
              "value": "8.53",
              "source": "/two-hours-on-battery/data/supplement-runs/tuned-cpu-laptop-m6.json derived.decode_tok_s.median 8.529 (num_thread 20)"
            },
            {
              "value": "13.05",
              "source": "/two-hours-on-battery/data/supplement-runs/tuned-cpu-laptop-m1.json derived.decode_tok_s.median 13.052 (num_thread 20)"
            }
          ]
        }
      ]
    }
  ],
  "source": "The battery page's kit: runs/cpu-{server,laptop,mini}-m5.json beside -m4, and -m6 beside -m1, derived.decode_tok_s.median; the third panel from supplement-runs/tuned-cpu-laptop-m5, -m4, -m6 and -m1.json (num_thread 20). Exhibit 105's kit: two-channel-desktop/summary.json arms[0] and arms[1].decode_median. Table T1 and its note."
}