Ergebnisse aktualisiert (20 Läufe)

app.js

@@ -1,36 +1,54 @@
 (function () {
   "use strict";
 
+  // Ansicht je Tab merken: Laufen Messungen, lädt sich die Seite alle 30 s neu (meta refresh). Sortierung, Filter
+  // und Vergleichsauswahl sollen das überstehen. Nur Bequemlichkeit – ohne Speicher startet die Seite wie geliefert.
+  function viewKey(name) { return "bench.view:" + location.pathname + location.search + ":" + name; }
+  function loadView(name) {
+    try { return JSON.parse(window.sessionStorage.getItem(viewKey(name))); } catch (e) { return null; }
+  }
+  function saveView(name, value) {
+    try { window.sessionStorage.setItem(viewKey(name), JSON.stringify(value)); } catch (e) { /* ohne Speicher */ }
+  }
+
   // Tabellen sortieren (Klick auf th[data-sort]); leere Werte stehen immer am Ende.
-  document.querySelectorAll("table.sortable").forEach(function (table) {
+  document.querySelectorAll("table.sortable").forEach(function (table, tableIndex) {
+    var name = "sort:" + (table.id || tableIndex);
+    function sortBy(th, dir) {
+      var index = Array.prototype.indexOf.call(th.parentNode.children, th);
+      var numeric = th.dataset.sort === "num";
+      table.querySelectorAll("th[aria-sort]").forEach(function (other) { other.removeAttribute("aria-sort"); });
+      th.setAttribute("aria-sort", dir);
+      var body = table.tBodies[0];
+      var rows = Array.prototype.slice.call(body.rows);
+      function valueOf(row) {
+        var cell = row.cells[index];
+        return cell.dataset.value !== undefined ? cell.dataset.value : cell.textContent.trim();
+      }
+      rows.sort(function (a, b) {
+        var va = valueOf(a), vb = valueOf(b);
+        if (numeric) {
+          var na = va === "" ? null : parseFloat(va);
+          var nb = vb === "" ? null : parseFloat(vb);
+          if (na === null && nb === null) return 0;
+          if (na === null) return 1;
+          if (nb === null) return -1;
+          return dir === "ascending" ? na - nb : nb - na;
+        }
+        return dir === "ascending" ? va.localeCompare(vb, "de") : vb.localeCompare(va, "de");
+      });
+      rows.forEach(function (row) { body.appendChild(row); });
+      return index;
+    }
     table.querySelectorAll("th[data-sort]").forEach(function (th) {
       th.addEventListener("click", function () {
-        var index = Array.prototype.indexOf.call(th.parentNode.children, th);
-        var numeric = th.dataset.sort === "num";
         var dir = th.getAttribute("aria-sort") === "descending" ? "ascending" : "descending";
-        table.querySelectorAll("th[aria-sort]").forEach(function (other) { other.removeAttribute("aria-sort"); });
-        th.setAttribute("aria-sort", dir);
-        var body = table.tBodies[0];
-        var rows = Array.prototype.slice.call(body.rows);
-        function valueOf(row) {
-          var cell = row.cells[index];
-          return cell.dataset.value !== undefined ? cell.dataset.value : cell.textContent.trim();
-        }
-        rows.sort(function (a, b) {
-          var va = valueOf(a), vb = valueOf(b);
-          if (numeric) {
-            var na = va === "" ? null : parseFloat(va);
-            var nb = vb === "" ? null : parseFloat(vb);
-            if (na === null && nb === null) return 0;
-            if (na === null) return 1;
-            if (nb === null) return -1;
-            return dir === "ascending" ? na - nb : nb - na;
-          }
-          return dir === "ascending" ? va.localeCompare(vb, "de") : vb.localeCompare(va, "de");
-        });
-        rows.forEach(function (row) { body.appendChild(row); });
+        saveView(name, { column: sortBy(th, dir), dir: dir });
       });
     });
+    var saved = loadView(name);
+    var header = saved && table.tHead ? table.tHead.rows[0].cells[saved.column] : null;
+    if (header && header.dataset.sort) sortBy(header, saved.dir === "ascending" ? "ascending" : "descending");
   });
 
   // Zeilen nach Reasoning-Modus filtern. Die Schalter stehen nur im HTML, wenn eine Tabelle mehrere Modi
@@ -38,33 +56,46 @@
   document.querySelectorAll("[data-filter-table]").forEach(function (group) {
     var table = document.getElementById(group.dataset.filterTable);
     if (!table || !table.tBodies[0]) return;
+    var name = "filter:" + group.dataset.filterTable;
     var buttons = Array.prototype.slice.call(group.querySelectorAll("[data-reasoning-filter]"));
-    buttons.forEach(function (button) {
-      button.addEventListener("click", function () {
-        var wanted = button.dataset.reasoningFilter;
-        buttons.forEach(function (other) {
-          var picked = other === button;
-          other.classList.toggle("active", picked);
-          other.setAttribute("aria-pressed", picked ? "true" : "false");
-        });
-        Array.prototype.slice.call(table.tBodies[0].rows).forEach(function (row) {
-          row.hidden = wanted !== "all" && (row.dataset.reasoning || "default") !== wanted;
-        });
+    function apply(button) {
+      var wanted = button.dataset.reasoningFilter;
+      buttons.forEach(function (other) {
+        var picked = other === button;
+        other.classList.toggle("active", picked);
+        other.setAttribute("aria-pressed", picked ? "true" : "false");
+      });
+      Array.prototype.slice.call(table.tBodies[0].rows).forEach(function (row) {
+        row.hidden = wanted !== "all" && (row.dataset.reasoning || "default") !== wanted;
       });
+      return wanted;
+    }
+    buttons.forEach(function (button) {
+      button.addEventListener("click", function () { saveView(name, apply(button)); });
     });
+    var saved = loadView(name);
+    var match = buttons.filter(function (b) { return b.dataset.reasoningFilter === saved; })[0];
+    if (match) apply(match);
   });
 
   // Läufe für den Vergleich auswählen (mindestens 2).
   var picks = document.querySelectorAll(".compare-pick");
   var button = document.getElementById("compare-button");
   var hidden = document.getElementById("compare-runs");
-  picks.forEach(function (box) {
-    box.addEventListener("change", function () {
-      var ids = Array.prototype.filter.call(picks, function (p) { return p.checked; }).map(function (p) { return p.value; });
-      hidden.value = ids.join(",");
-      button.disabled = ids.length < 2;
+  function updatePicks() {
+    var ids = Array.prototype.filter.call(picks, function (p) { return p.checked; }).map(function (p) { return p.value; });
+    hidden.value = ids.join(",");
+    button.disabled = ids.length < 2;
+    return ids;
+  }
+  if (picks.length && button && hidden) {
+    var savedPicks = loadView("picks") || [];
+    picks.forEach(function (box) {
+      if (savedPicks.indexOf(box.value) >= 0) box.checked = true;
+      box.addEventListener("change", function () { saveView("picks", updatePicks()); });
     });
-  });
+    updatePicks();
+  }
 
   // Bestätigung vor gefährlichen Aktionen (Formulare mit data-confirm).
   document.querySelectorAll("form[data-confirm]").forEach(function (form) {
@@ -143,6 +174,14 @@
   function rememberCategory(key) {
     try { window.localStorage.setItem(CATEGORY_KEY, key); } catch (e) { /* ohne Speicher */ }
   }
+  // Ebenso die x-Achse des Streudiagramms: Tokens/s ("tps") oder Rechenzeit der Suite ("busy").
+  var X_KEY = "bench.chartX";
+  function rememberedX() {
+    try { return window.localStorage.getItem(X_KEY) === "busy" ? "busy" : "tps"; } catch (e) { return "tps"; }
+  }
+  function rememberX(mode) {
+    try { window.localStorage.setItem(X_KEY, mode); } catch (e) { /* ohne Speicher */ }
+  }
 
   // Übersicht (Webseite und Space): je .chart-set ein Balken- und ein Streudiagramm, umschaltbar nach Kategorie.
   document.querySelectorAll(".chart-set").forEach(function (set) {
@@ -152,10 +191,20 @@
     var texts = data.texts;
     var limits = data.thresholds || {};
     var buttons = Array.prototype.slice.call(set.querySelectorAll("[data-chart-category]"));
+    var xButtons = Array.prototype.slice.call(set.querySelectorAll("[data-chart-x]"));
+    var hints = Array.prototype.slice.call(set.querySelectorAll("[data-chart-hint]"));
+    var xMode = xButtons.length ? rememberedX() : "tps";
+    var currentKey = "overall";
     var barBox = set.querySelector('[data-chart-box="bar"]');
     var barTitle = set.querySelector('[data-chart-title="bar"]');
     var scatterTitle = set.querySelector('[data-chart-title="scatter"]');
-    var barChart = null, scatterChart = null;
+    var barChart = null, scatterChart = null, hoverBar = null, hoverScatter = null;
+    var barCanvas = set.querySelector('[data-chart="bar"]');
+    var scatterCanvas = set.querySelector('[data-chart="scatter"]');
+    barCanvas.addEventListener("mousemove", function (event) { if (hoverBar) hoverBar(event); });
+    barCanvas.addEventListener("mouseleave", function () { if (hoverBar) hoverBar(null); });
+    scatterCanvas.addEventListener("mousemove", function (event) { if (hoverScatter) hoverScatter(event); });
+    scatterCanvas.addEventListener("mouseleave", function () { if (hoverScatter) hoverScatter(null); });
 
     function titleOf(key) {
       if (key === "overall") return texts.overall_title;
@@ -163,22 +212,37 @@
       return match ? match.label : key;
     }
 
+    function busyText(r) {
+      return (r.busy_estimated ? "≈ " : "") + r.busy_h.toFixed(1) + " " + texts.busy_unit;
+    }
+
     function show(key) {
+      currentKey = key;
       var title = titleOf(key);
+      var busyMode = xMode === "busy";
       buttons.forEach(function (b) {
         var picked = b.dataset.chartCategory === key;
         b.classList.toggle("active", picked);
         b.setAttribute("aria-pressed", picked ? "true" : "false");
       });
+      xButtons.forEach(function (b) {
+        var picked = b.dataset.chartX === xMode;
+        b.classList.toggle("active", picked);
+        b.setAttribute("aria-pressed", picked ? "true" : "false");
+      });
+      hints.forEach(function (h) { h.hidden = h.dataset.chartHint !== xMode; });
       if (barTitle) barTitle.textContent = title;
-      if (scatterTitle) scatterTitle.textContent = title + texts.vs;
+      if (scatterTitle) scatterTitle.textContent = title + (busyMode ? texts.vs_busy : texts.vs);
       var bars = data.runs.filter(function (r) { return r.scores[key] !== null && r.scores[key] !== undefined; })
         .sort(function (a, b) { return b.scores[key] - a.scores[key]; });
-      var points = bars.filter(function (r) { return r.tps !== null; })
-        .map(function (r) { return { label: r.label, x: r.tps, y: r.scores[key], source: r.tps_source }; });
-      // Hohl: Geschwindigkeit nur aus dem Speed-Test, der Lauf lieferte keine Werte.
-      var fills = points.map(function (p) { return p.source === "test" ? surface : series[0]; });
-      var edges = points.map(function (p) { return p.source === "test" ? series[0] : surface; });
+      // Hohl: bei Tokens/s nur der Speed-Test-Wert (der Lauf lieferte keine), bei Rechenzeit ein geschätzter Wert.
+      var points = busyMode
+        ? bars.filter(function (r) { return r.busy_h !== null && r.busy_h !== undefined; })
+          .map(function (r) { return { label: r.label, x: r.busy_h, y: r.scores[key], hollow: r.busy_estimated }; })
+        : bars.filter(function (r) { return r.tps !== null; })
+          .map(function (r) { return { label: r.label, x: r.tps, y: r.scores[key], hollow: r.tps_source === "test" }; });
+      var fills = points.map(function (p) { return p.hollow ? surface : series[0]; });
+      var edges = points.map(function (p) { return p.hollow ? series[0] : surface; });
 
       // Balken der Modelle im Wunschbereich grün – dieselbe Regel wie das Rechteck im Streudiagramm.
       var minTps = limits.min_gen_tps == null ? null : limits.min_gen_tps;
@@ -189,9 +253,41 @@
         return minScore === null || r.scores[key] >= minScore;
       }
 
+      // Hover verknüpft beide Diagramme: Balken oder Modellname markiert den Punkt desselben Laufs und umgekehrt.
+      function mark(chart, index) {
+        var current = chart.getActiveElements()[0];
+        if (current ? current.index === index : index < 0) return;
+        var active = index < 0 ? [] : [{ datasetIndex: 0, index: index }];
+        var el = index < 0 ? null : chart.getDatasetMeta(0).data[index];
+        chart.setActiveElements(active);
+        chart.tooltip.setActiveElements(active, el ? { x: el.x, y: el.y } : { x: 0, y: 0 });
+        chart.update();
+      }
+      function indexOf(list, label) {
+        return label === null ? -1 : list.findIndex(function (r) { return r.label === label; });
+      }
+      // Eigene Listener statt onHover: Chart.js meldet nichts außerhalb des Diagrammbereichs (Modellnamen, Verlassen).
+      hoverBar = function (event) {
+        if (!scatterChart) return;
+        if (!event) return mark(scatterChart, -1);
+        var rect = barChart.canvas.getBoundingClientRect();
+        var x = event.clientX - rect.left, y = event.clientY - rect.top;
+        var area = barChart.chartArea;
+        var index = y < area.top || y > area.bottom ? -1 : Math.round(barChart.scales.y.getValueForPixel(y));
+        if (index < 0 || index >= bars.length) return mark(scatterChart, -1);
+        // Über den Namen zeigt auch der Balken seinen Tooltip, sonst bliebe der des zuletzt berührten Balkens stehen.
+        if (x < area.left) mark(barChart, index);
+        mark(scatterChart, indexOf(points, bars[index].label));
+      };
+      hoverScatter = function (event) {
+        if (!barChart) return;
+        var hit = event ? scatterChart.getElementsAtEventForMode(event, "nearest", { intersect: true }, false)[0] : null;
+        mark(barChart, indexOf(bars, hit ? points[hit.index].label : null));
+      };
+
       if (barChart) barChart.destroy();
       if (barBox) barBox.style.height = (70 + 30 * Math.max(bars.length, 1)) + "px";
-      barChart = new Chart(set.querySelector('[data-chart="bar"]'), {
+      barChart = new Chart(barCanvas, {
         type: "bar",
         data: {
           labels: bars.map(function (r) { return r.label; }),
@@ -201,6 +297,8 @@
             backgroundColor: bars.map(function (r) { return inZone(r) ? series[2] : series[0]; }),
             borderRadius: 4,
             borderSkipped: "start",
+            hoverBorderColor: ink,
+            hoverBorderWidth: 2,
             barPercentage: 0.8,
             categoryPercentage: 0.9,
             maxBarThickness: 18
@@ -214,16 +312,27 @@
           scales: { x: axis(texts.score_axis, { min: 0, max: 100 }), y: axis(null, { grid: { display: false } }) },
           plugins: {
             legend: { display: false },
-            tooltip: { callbacks: { label: function (c) { return title + ": " + c.parsed.x.toFixed(1); } } }
+            tooltip: {
+              callbacks: {
+                label: function (c) { return title + ": " + c.parsed.x.toFixed(1); },
+                afterLabel: function (c) {
+                  var r = bars[c.dataIndex], lines = [];
+                  if (r.tps !== null) lines.push(r.tps.toFixed(1) + " t/s");
+                  if (r.busy_h !== null && r.busy_h !== undefined) lines.push(busyText(r));
+                  return lines;
+                }
+              }
+            }
           }
         },
         plugins: [barValueLabels]
       });
 
       if (scatterChart) scatterChart.destroy();
-      var minX = limits.min_gen_tps == null ? null : limits.min_gen_tps;
-      var minY = limits.min_score == null ? null : limits.min_score;
-      scatterChart = new Chart(set.querySelector('[data-chart="scatter"]'), {
+      // Der Wunschbereich hängt an min. Tokens/s und gilt deshalb nur auf der Tokens/s-Achse.
+      var minX = busyMode || limits.min_gen_tps == null ? null : limits.min_gen_tps;
+      var minY = busyMode || limits.min_score == null ? null : limits.min_score;
+      scatterChart = new Chart(scatterCanvas, {
         type: "scatter",
         data: {
           datasets: [{
@@ -233,7 +342,9 @@
             borderColor: edges,
             borderWidth: 2,
             pointRadius: 5,
-            pointHoverRadius: 7,
+            pointHoverRadius: 8,
+            pointHoverBorderColor: ink,
+            pointHoverBorderWidth: 2,
             pointHitRadius: 12
           }]
         },
@@ -242,7 +353,10 @@
           animation: false,
           scales: {
             // Der Wunschbereich bleibt sichtbar, auch wenn alle Modelle langsamer sind.
-            x: axis(texts.tps_axis, { beginAtZero: true, suggestedMax: minX === null ? undefined : minX * 1.1 }),
+            // Rechenzeit gespiegelt: rechts steht wie bei Tokens/s das schnellere Modell.
+            x: busyMode
+              ? axis(texts.busy_axis, { beginAtZero: true, reverse: true })
+              : axis(texts.tps_axis, { beginAtZero: true, suggestedMax: minX === null ? undefined : minX * 1.1 }),
             y: axis(title, { min: 0, max: 100 })
           },
           plugins: {
@@ -251,8 +365,10 @@
             tooltip: {
               callbacks: {
                 label: function (c) {
-                  var source = c.raw.source === "test" ? texts.source_test : texts.source_run;
-                  return c.raw.label + ": " + c.raw.y.toFixed(1) + " " + texts.points + ", " + c.raw.x.toFixed(1) + " t/s (" + source + ")";
+                  var head = c.raw.label + ": " + c.raw.y.toFixed(1) + " " + texts.points + ", ";
+                  if (busyMode) return head + c.raw.x.toFixed(1) + " " + texts.busy_unit + (c.raw.hollow ? " (" + texts.estimated + ")" : "");
+                  var source = c.raw.hollow ? texts.source_test : texts.source_run;
+                  return head + c.raw.x.toFixed(1) + " t/s (" + source + ")";
                 }
               }
             }
@@ -268,6 +384,13 @@
         rememberCategory(b.dataset.chartCategory);
       });
     });
+    xButtons.forEach(function (b) {
+      b.addEventListener("click", function () {
+        xMode = b.dataset.chartX;
+        rememberX(xMode);
+        show(currentKey);
+      });
+    });
     var start = rememberedCategory();
     var known = data.categories.some(function (c) { return c.key === start; });
     show(known ? start : "overall");
@@ -288,6 +411,8 @@
             borderWidth: { top: 1, bottom: 1, right: 2 },
             borderRadius: 4,
             borderSkipped: "start",
+            hoverBorderColor: ink,
+            hoverBorderWidth: 2,
             barPercentage: 0.9,
             categoryPercentage: 0.8,
             maxBarThickness: 16

index.html

@@ -29,9 +29,46 @@
 <thead><tr>
   <th data-sort="text">Model</th>
   <th data-sort="num" class="num">Overall</th>
-<th data-sort="num" class="num">Knowledge &amp; reasoning</th><th data-sort="num" class="num">Math</th><th data-sort="num" class="num">Instruction following</th><th data-sort="num" class="num">Code</th><th data-sort="num" class="num">German</th><th data-sort="num" class="num">gen t/s</th><th data-sort="num" class="num" title="Generation per request measured during the benchmark run; gen t/s before it comes from the synthetic speed test">gen t/s run</th><th data-sort="num" class="num">pp t/s 4k</th><th data-sort="num" class="num">TTFT 4k ms</th>  <th data-sort="text">Date</th>
+<th data-sort="num" class="num">Knowledge &amp; reasoning</th><th data-sort="num" class="num">Math</th><th data-sort="num" class="num">Instruction following</th><th data-sort="num" class="num">Code</th><th data-sort="num" class="num">German</th><th data-sort="num" class="num">gen t/s</th><th data-sort="num" class="num" title="Generation per request measured during the benchmark run; gen t/s before it comes from the synthetic speed test">gen t/s run</th><th data-sort="num" class="num">pp t/s 4k</th><th data-sort="num" class="num">TTFT 4k ms</th>  <th data-sort="num" class="num" title="Time during which the run's requests were in flight for the whole suite, excluding pauses and reruns – how long a model really takes for the tasks: speed times answer length">Compute time [h]</th>
+  <th data-sort="text">Date</th>
 </tr></thead>
 <tbody>
+<tr data-reasoning="on">
+  <td data-value="qwen3.8-flash-next-ple4">
+    <a href="runs/28.html" class="strong">qwen3.8-flash-next-ple4</a>
+    <div class="sub">W4A16 · Reasoning on · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="76.40898610161769">76.4</td>
+  <td class="num" data-value="79.83030303030303" title="mmlu_pro: 89.3 ± 1.8 (n=300)&#10;gpqa_diamond_gen: 75.8 ± 3.1 (n=198)&#10;musr_gen: 74.4 ± 2.8 (n=250)&#10;">79.8</td>
+  <td class="num" data-value="66.0" title="omni_math_gen: 66.0 ± 3.0 (n=250)&#10;">66.0</td>
+  <td class="num" data-value="77.28822055137846" title="ifbench: 71.3 ± 2.6 (n=300)&#10;ifeval: 95.3 ± 1.7 (n=150)&#10;">77.3</td>
+  <td class="num" data-value="78.0" title="livecodebench_gen: 78.0 ± 2.9 (n=200)&#10;">78.0</td>
+  <td class="num" data-value="80.92640692640693" title="mmlu_prox_de: 86.0 ± 2.0 (n=300)&#10;include_de_gen: 70.8 ± 3.7 (n=154)&#10;">80.9</td>
+  <td class="num" data-value="25.15751834546852">25</td>
+  <td class="num" data-value="20.13654359300451" title="overall 60.9 t/s with 3 parallel requests, from 9 of 9 tasks">20</td>
+  <td class="num" data-value="1952.3318729559924">1952</td>
+  <td class="num" data-value="2001.7088560271077">2002</td>
+<td class="num" data-value="43.52194444444444">43.5</td>
+  <td data-value="2026-09-21">2026-09-21</td>
+</tr>
+<tr data-reasoning="on">
+  <td data-value="qwen3.8-flash-next-abliterated">
+    <a href="runs/34.html" class="strong">qwen3.8-flash-next-abliterated</a>
+    <div class="sub">iQ3-S · Reasoning on · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="76.21910787068683">76.2</td>
+  <td class="num strong" data-value="82.04915824915825" title="mmlu_pro: 89.0 ± 1.8 (n=300)&#10;gpqa_diamond_gen: 74.7 ± 3.1 (n=198)&#10;musr_gen: 82.4 ± 2.4 (n=250)&#10;">82.0</td>
+  <td class="num" data-value="65.60000000000001" title="omni_math_gen: 65.6 ± 3.0 (n=250)&#10;">65.6</td>
+  <td class="num" data-value="76.79197994987469" title="ifbench: 71.3 ± 2.6 (n=300)&#10;ifeval: 93.3 ± 2.0 (n=150)&#10;">76.8</td>
+  <td class="num" data-value="75.5" title="livecodebench_gen: 75.5 ± 3.0 (n=200)&#10;">75.5</td>
+  <td class="num" data-value="81.15440115440116" title="mmlu_prox_de: 86.7 ± 2.0 (n=300)&#10;include_de_gen: 70.1 ± 3.7 (n=154)&#10;">81.2</td>
+  <td class="num" data-value="34.80493986372224">35</td>
+  <td class="num" data-value="19.49242514051305" title="overall 59.6 t/s with 3 parallel requests, from 9 of 9 tasks">19</td>
+  <td class="num" data-value="575.3031643138398">575</td>
+  <td class="num" data-value="6792.9402138106525">6793</td>
+<td class="num" data-value="46.07227777777778">46.1</td>
+  <td data-value="2026-10-11">2026-10-11</td>
+</tr>
 <tr data-reasoning="on">
   <td data-value="gpt-oss-120b">
     <a href="runs/18.html" class="strong">gpt-oss-120b</a>
@@ -40,15 +77,34 @@
   <td class="num strong" data-value="73.08253056884635">73.1</td>
   <td class="num" data-value="68.52727272727272" title="mmlu_pro: 71.0 ± 2.6 (n=300)&#10;gpqa_diamond_gen: 68.2 ± 3.3 (n=198)&#10;musr_gen: 66.4 ± 3.0 (n=250)&#10;">68.5</td>
   <td class="num" data-value="70.39999999999999" title="omni_math_gen: 70.4 ± 2.9 (n=250)&#10;">70.4</td>
-  <td class="num strong" data-value="66.1203007518797" title="ifbench: 60.0 ± 2.8 (n=300)&#10;ifeval: 84.7 ± 3.0 (n=150)&#10;">66.1</td>
+  <td class="num" data-value="66.1203007518797" title="ifbench: 60.0 ± 2.8 (n=300)&#10;ifeval: 84.7 ± 3.0 (n=150)&#10;">66.1</td>
   <td class="num strong" data-value="83.0" title="livecodebench_gen: 83.0 ± 2.7 (n=200)&#10;">83.0</td>
   <td class="num" data-value="77.36507936507935" title="mmlu_prox_de: 80.3 ± 2.3 (n=300)&#10;include_de_gen: 71.4 ± 3.7 (n=154)&#10;">77.4</td>
   <td class="num" data-value="">–</td>
   <td class="num" data-value="19.638176195686693" title="overall 58.4 t/s with 3 parallel requests, from 9 of 9 tasks">20</td>
   <td class="num" data-value="">–</td>
   <td class="num" data-value="">–</td>
+<td class="num" data-value="15.209277777777777">15.2</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
+<tr data-reasoning="off">
+  <td data-value="qwen3.8-flash-next-abliterated">
+    <a href="runs/35.html" class="strong">qwen3.8-flash-next-abliterated</a>
+    <div class="sub">iQ3-S · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="72.82945950735424">72.8</td>
+  <td class="num" data-value="77.68148148148148" title="mmlu_pro: 83.7 ± 2.1 (n=300)&#10;gpqa_diamond_gen: 77.8 ± 3.0 (n=198)&#10;musr_gen: 71.6 ± 2.9 (n=250)&#10;">77.7</td>
+  <td class="num" data-value="71.2" title="omni_math_gen: 71.2 ± 2.9 (n=250)&#10;">71.2</td>
+  <td class="num" data-value="53.578947368421055" title="ifbench: 42.0 ± 2.9 (n=300)&#10;ifeval: 88.7 ± 2.6 (n=150)&#10;">53.6</td>
+  <td class="num" data-value="81.0" title="livecodebench_gen: 81.0 ± 2.8 (n=200)&#10;">81.0</td>
+  <td class="num" data-value="80.68686868686869" title="mmlu_prox_de: 84.7 ± 2.1 (n=300)&#10;include_de_gen: 72.7 ± 3.6 (n=154)&#10;">80.7</td>
+  <td class="num" data-value="38.111744479072115">38</td>
+  <td class="num" data-value="20.913074181761658" title="overall 64.0 t/s with 3 parallel requests, from 9 of 9 tasks">21</td>
+  <td class="num" data-value="658.2705125331207">658</td>
+  <td class="num" data-value="6112.988389097154">6113</td>
+<td class="num" data-value="13.837194444444444">13.8</td>
+  <td data-value="2026-10-12">2026-10-12</td>
+</tr>
 <tr data-reasoning="off">
   <td data-value="qwen3.8-flash-next-abliterated (pool)">
     <a href="runs/32.html" class="strong">qwen3.8-flash-next-abliterated (pool)</a>
@@ -64,6 +120,7 @@
   <td class="num" data-value="37.23226449422559" title="overall 113.3 t/s with 3 parallel requests, from 9 of 9 tasks">37</td>
   <td class="num" data-value="2569.0663350056934">2569</td>
   <td class="num" data-value="1566.3277919957181">1566</td>
+<td class="num" data-value="8.065444444444445">8.1</td>
   <td data-value="2026-09-27">2026-09-27</td>
 </tr>
 <tr data-reasoning="off">
@@ -72,7 +129,7 @@
     <div class="sub">W4A16 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM</div>
   </td>
   <td class="num strong" data-value="71.45393635604162">71.5</td>
-  <td class="num strong" data-value="77.86868686868686" title="mmlu_pro: 86.3 ± 2.0 (n=300)&#10;gpqa_diamond_gen: 77.3 ± 3.0 (n=198)&#10;musr_gen: 70.0 ± 2.9 (n=250)&#10;">77.9</td>
+  <td class="num" data-value="77.86868686868686" title="mmlu_pro: 86.3 ± 2.0 (n=300)&#10;gpqa_diamond_gen: 77.3 ± 3.0 (n=198)&#10;musr_gen: 70.0 ± 2.9 (n=250)&#10;">77.9</td>
   <td class="num strong" data-value="73.2" title="omni_math_gen: 73.2 ± 2.8 (n=250)&#10;">73.2</td>
   <td class="num" data-value="49.56390977443609" title="ifbench: 36.0 ± 2.8 (n=300)&#10;ifeval: 90.7 ± 2.4 (n=150)&#10;">49.6</td>
   <td class="num" data-value="75.5" title="livecodebench_gen: 75.5 ± 3.0 (n=200)&#10;">75.5</td>
@@ -81,6 +138,7 @@
   <td class="num" data-value="21.59310838133867" title="overall 64.4 t/s with 3 parallel requests, from 5 of 9 tasks">22</td>
   <td class="num" data-value="2041.4640201348918">2041</td>
   <td class="num" data-value="1973.583644023165">1974</td>
+<td class="num" data-value="14.038722222222223" title="estimated: tasks without a request log count with their wall-clock duration">≈ 14.0</td>
   <td data-value="2026-09-17">2026-09-17</td>
 </tr>
 <tr data-reasoning="off">
@@ -98,6 +156,7 @@
   <td class="num" data-value="13.757625586585043" title="overall 41.2 t/s with 3 parallel requests, from 5 of 9 tasks">14</td>
   <td class="num" data-value="1359.0489942173106">1359</td>
   <td class="num" data-value="2952.064286917448">2952</td>
+<td class="num" data-value="12.862305555555555" title="estimated: tasks without a request log count with their wall-clock duration">≈ 12.9</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
 <tr data-reasoning="off">
@@ -115,8 +174,27 @@
   <td class="num" data-value="13.819760405181762" title="overall 41.0 t/s with 3 parallel requests, from 5 of 9 tasks">14</td>
   <td class="num" data-value="1358.107397713914">1358</td>
   <td class="num" data-value="2954.110997961834">2954</td>
+<td class="num" data-value="12.531166666666667" title="estimated: tasks without a request log count with their wall-clock duration">≈ 12.5</td>
   <td data-value="2026-09-19">2026-09-19</td>
 </tr>
+<tr data-reasoning="on">
+  <td data-value="lfm2.5-24b-a2b-0">
+    <a href="runs/36.html" class="strong">lfm2.5-24b-a2b-0</a>
+    <div class="sub">Q8 · Reasoning on · direct �� NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="65.71215918584339">65.7</td>
+  <td class="num" data-value="68.05252525252526" title="mmlu_pro: 76.7 ± 2.4 (n=300)&#10;gpqa_diamond_gen: 59.1 ± 3.5 (n=198)&#10;musr_gen: 68.4 ± 2.9 (n=250)&#10;">68.1</td>
+  <td class="num" data-value="65.2" title="omni_math_gen: 65.2 ± 3.0 (n=250)&#10;">65.2</td>
+  <td class="num strong" data-value="79.8796992481203" title="ifbench: 75.0 ± 2.5 (n=300)&#10;ifeval: 94.7 ± 1.8 (n=150)&#10;">79.9</td>
+  <td class="num" data-value="44.0" title="livecodebench_gen: 44.0 ± 3.5 (n=200)&#10;">44.0</td>
+  <td class="num" data-value="71.42857142857143" title="mmlu_prox_de: 75.0 ± 2.5 (n=300)&#10;include_de_gen: 64.3 ± 3.9 (n=154)&#10;">71.4</td>
+  <td class="num" data-value="77.10114095123421">77</td>
+  <td class="num" data-value="34.40430548575067" title="overall 89.1 t/s with 3 parallel requests, from 9 of 9 tasks">34</td>
+  <td class="num" data-value="2775.1919328454383">2775</td>
+  <td class="num" data-value="1451.0708078742027">1451</td>
+<td class="num" data-value="20.664805555555557">20.7</td>
+  <td data-value="2026-10-20">2026-10-20</td>
+</tr>
 <tr data-reasoning="off">
   <td data-value="qwen3-coder-next">
     <a href="runs/30.html" class="strong">qwen3-coder-next</a>
@@ -132,6 +210,7 @@
   <td class="num" data-value="23.80832467007433" title="overall 59.7 t/s with 3 parallel requests, from 9 of 9 tasks">24</td>
   <td class="num" data-value="987.8475690816915">988</td>
   <td class="num" data-value="4077.5521710747853">4078</td>
+<td class="num" data-value="16.2905">16.3</td>
   <td data-value="2026-09-23">2026-09-23</td>
 </tr>
 <tr data-reasoning="off">
@@ -149,6 +228,7 @@
   <td class="num" data-value="22.383948355383822" title="overall 67.6 t/s with 3 parallel requests, from 5 of 9 tasks">22</td>
   <td class="num" data-value="2550.5619640606633">2551</td>
   <td class="num" data-value="1577.6915270835161">1578</td>
+<td class="num" data-value="12.367083333333333" title="estimated: tasks without a request log count with their wall-clock duration">≈ 12.4</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
 <tr data-reasoning="off">
@@ -166,6 +246,7 @@
   <td class="num" data-value="31.562617796089718" title="overall 94.1 t/s with 3 parallel requests, from 9 of 9 tasks">32</td>
   <td class="num" data-value="5539.414736774487">5539</td>
   <td class="num" data-value="726.4305330463685">726</td>
+<td class="num" data-value="11.385472222222221">11.4</td>
   <td data-value="2026-09-21">2026-09-21</td>
 </tr>
 <tr data-reasoning="off">
@@ -183,6 +264,7 @@
   <td class="num" data-value="21.942627395574263" title="overall 67.0 t/s with 3 parallel requests, from 9 of 9 tasks">22</td>
   <td class="num" data-value="2472.0116600865786">2472</td>
   <td class="num" data-value="1627.824036986567">1628</td>
+<td class="num" data-value="12.698694444444444">12.7</td>
   <td data-value="2026-09-20">2026-09-20</td>
 </tr>
 <tr data-reasoning="off">
@@ -200,6 +282,7 @@
   <td class="num" data-value="19.303084023060574" title="overall 58.5 t/s with 3 parallel requests, from 9 of 9 tasks">19</td>
   <td class="num" data-value="898.1292759650371">898</td>
   <td class="num" data-value="4480.424041044898">4480</td>
+<td class="num" data-value="17.456333333333333">17.5</td>
   <td data-value="2026-09-21">2026-09-21</td>
 </tr>
 <tr data-reasoning="off">
@@ -217,8 +300,27 @@
   <td class="num" data-value="21.7770272387994" title="overall 70.2 t/s with 3 parallel requests, from 9 of 9 tasks">22</td>
   <td class="num" data-value="2177.560718839664">2178</td>
   <td class="num" data-value="1847.9392860026564">1848</td>
+<td class="num" data-value="7.183694444444445">7.2</td>
   <td data-value="2026-09-26">2026-09-26</td>
 </tr>
+<tr data-reasoning="off">
+  <td data-value="nemotron-3.5-super-vl-120b-a12b-vllm-lilicorr">
+    <a href="runs/37.html" class="strong">nemotron-3.5-super-vl-120b-a12b-vllm-lilicorr</a>
+    <div class="sub">NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="57.795315055315065">57.8</td>
+  <td class="num" data-value="71.31784511784512" title="mmlu_pro: 81.7 ± 2.2 (n=300)&#10;gpqa_diamond_gen: 68.7 ± 3.3 (n=198)&#10;musr_gen: 63.6 ± 3.0 (n=250)&#10;">71.3</td>
+  <td class="num" data-value="28.000000000000004" title="omni_math_gen: 28.0 ± 2.8 (n=250)&#10;">28.0</td>
+  <td class="num" data-value="53.23809523809524" title="ifbench: 40.7 ± 2.8 (n=300)&#10;ifeval: 91.3 ± 2.3 (n=150)&#10;">53.2</td>
+  <td class="num" data-value="59.5" title="livecodebench_gen: 59.5 ± 3.5 (n=200)&#10;">59.5</td>
+  <td class="num" data-value="76.92063492063492" title="mmlu_prox_de: 79.7 ± 2.3 (n=300)&#10;include_de_gen: 71.4 ± 3.7 (n=154)&#10;">76.9</td>
+  <td class="num" data-value="27.497155870102606">27</td>
+  <td class="num" data-value="25.267108590550038" title="overall 78.0 t/s with 3 parallel requests, from 9 of 9 tasks">25</td>
+  <td class="num" data-value="2878.373067788896">2878</td>
+  <td class="num" data-value="1392.453273292631">1392</td>
+<td class="num" data-value="4.086833333333334">4.1</td>
+  <td data-value="2026-10-21">2026-10-21</td>
+</tr>
 <tr data-reasoning="off">
   <td data-value="gemma-4-12b-it-heretic">
     <a href="runs/17.html" class="strong">gemma-4-12b-it-heretic</a>
@@ -234,6 +336,7 @@
   <td class="num" data-value="21.121219019103503" title="overall 62.6 t/s with 3 parallel requests, from 5 of 9 tasks">21</td>
   <td class="num" data-value="5258.027383235864">5258</td>
   <td class="num" data-value="763.404164230451">763</td>
+<td class="num" data-value="13.155916666666668" title="estimated: tasks without a request log count with their wall-clock duration">≈ 13.2</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
 <tr data-reasoning="off">
@@ -251,6 +354,7 @@
   <td class="num" data-value="70.42131680616991" title="overall 213.6 t/s with 3 parallel requests, from 5 of 9 tasks">70</td>
   <td class="num" data-value="7236.849600555504">7237</td>
   <td class="num" data-value="554.1085170116276">554</td>
+<td class="num" data-value="4.412361111111111" title="estimated: tasks without a request log count with their wall-clock duration">≈ 4.4</td>
   <td data-value="2026-09-17">2026-09-17</td>
 </tr>
 <tr data-reasoning="off">
@@ -268,6 +372,7 @@
   <td class="num" data-value="35.229241139471675" title="overall 106.0 t/s with 3 parallel requests, from 5 of 9 tasks">35</td>
   <td class="num" data-value="6290.925546819461">6291</td>
   <td class="num" data-value="639.6515059750527">640</td>
+<td class="num" data-value="13.010333333333332" title="estimated: tasks without a request log count with their wall-clock duration">≈ 13.0</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
 <tr data-reasoning="off">
@@ -285,29 +390,32 @@
   <td class="num" data-value="34.0307978654491" title="overall 102.4 t/s with 3 parallel requests, from 9 of 9 tasks">34</td>
   <td class="num" data-value="3957.900821130251">3958</td>
   <td class="num" data-value="1022.5117260124534">1023</td>
+<td class="num" data-value="9.628555555555556">9.6</td>
   <td data-value="2026-09-23">2026-09-23</td>
 </tr>
 </tbody>
 </table>
 </div>
 <div class="chart-set">
-<script type="application/json" class="chart-data" id="charts-v3-quick">{"categories": [{"key": "overall", "label": "Overall"}, {"key": "knowledge", "label": "Knowledge \u0026 reasoning"}, {"key": "math", "label": "Math"}, {"key": "instructions", "label": "Instruction following"}, {"key": "code", "label": "Code"}, {"key": "german", "label": "German"}], "missing_tps": false, "reasoning_mark": "\ud83d\udcad", "runs": [{"label": "gpt-oss-120b \u00b7 \ud83d\udcad", "run_id": 18, "scores": {"code": 83.0, "german": 77.4, "instructions": 66.1, "knowledge": 68.5, "math": 70.4, "overall": 73.1}, "tps": 19.6, "tps_source": "run"}, {"label": "qwen3.8-flash-next-abliterated (pool)", "run_id": 32, "scores": {"code": 77.0, "german": 81.4, "instructions": 55.2, "knowledge": 77.2, "math": 72.0, "overall": 72.5}, "tps": 37.2, "tps_source": "run"}, {"label": "qwen3.8-flash-next-ple4", "run_id": 15, "scores": {"code": 75.5, "german": 81.1, "instructions": 49.6, "knowledge": 77.9, "math": 73.2, "overall": 71.5}, "tps": 21.6, "tps_source": "run"}, {"label": "gemma-4-12b-it", "run_id": 21, "scores": {"code": 75.0, "german": 74.5, "instructions": 51.3, "knowledge": 67.5, "math": 60.8, "overall": 65.8}, "tps": 13.8, "tps_source": "run"}, {"label": "gemma-4-12b-it-heretic", "run_id": 22, "scores": {"code": 75.0, "german": 76.5, "instructions": 50.0, "knowledge": 68.4, "math": 59.2, "overall": 65.8}, "tps": 13.8, "tps_source": "run"}, {"label": "qwen3-coder-next", "run_id": 30, "scores": {"code": 70.5, "german": 76.0, "instructions": 45.9, "knowledge": 71.3, "math": 61.6, "overall": 65.1}, "tps": 23.8, "tps_source": "run"}, {"label": "qwen3.8-27b-mtp", "run_id": 20, "scores": {"code": 65.0, "german": 76.3, "instructions": 49.5, "knowledge": 69.9, "math": 60.4, "overall": 64.2}, "tps": 22.4, "tps_source": "run"}, {"label": "occamy-1.0", "run_id": 27, "scores": {"code": 63.0, "german": 76.9, "instructions": 44.7, "knowledge": 73.1, "math": 62.4, "overall": 64.0}, "tps": 31.6, "tps_source": "run"}, {"label": "qwen3.8-27b-abliterated", "run_id": 23, "scores": {"code": 59.5, "german": 72.1, "instructions": 46.1, "knowledge": 71.5, "math": 58.0, "overall": 61.4}, "tps": 21.9, "tps_source": "run"}, {"label": "ternary-bonsai-2-27b", "run_id": 26, "scores": {"code": 56.5, "german": 74.5, "instructions": 46.3, "knowledge": 69.5, "math": 59.6, "overall": 61.3}, "tps": 19.3, "tps_source": "run"}, {"label": "mimo-v2.6-flash (pool)", "run_id": 31, "scores": {"code": 61.5, "german": 74.5, "instructions": 49.8, "knowledge": 69.7, "math": 41.6, "overall": 59.4}, "tps": 21.8, "tps_source": "run"}, {"label": "gemma-4-12b-it-heretic", "run_id": 17, "scores": {"code": 62.0, "german": 67.7, "instructions": 46.8, "knowledge": 63.7, "math": 48.4, "overall": 57.7}, "tps": 21.1, "tps_source": "run"}, {"label": "nemotron-3.5-lightning-dspark", "run_id": 16, "scores": {"code": 53.0, "german": 70.5, "instructions": 46.0, "knowledge": 55.8, "math": 40.4, "overall": 53.1}, "tps": 70.4, "tps_source": "run"}, {"label": "qwen3.5-9b-heretic-v2-mtp", "run_id": 19, "scores": {"code": 46.5, "german": 67.0, "instructions": 44.1, "knowledge": 65.1, "math": 35.6, "overall": 51.7}, "tps": 35.2, "tps_source": "run"}, {"label": "zgcm-1-7b", "run_id": 29, "scores": {"code": 23.0, "german": 24.8, "instructions": 44.3, "knowledge": 34.7, "math": 40.0, "overall": 33.4}, "tps": 34.0, "tps_source": "run"}], "texts": {"overall": "Overall", "overall_title": "Overall score", "points": "points", "score_axis": "Score (0\u2013100)", "source_run": "measured during the run", "source_test": "from the speed test", "tps_axis": "Generation (tokens/s)", "vs": " vs. generation (tokens/s)"}}</script>
+<script type="application/json" class="chart-data" id="charts-v3-quick">{"categories": [{"key": "overall", "label": "Overall"}, {"key": "knowledge", "label": "Knowledge \u0026 reasoning"}, {"key": "math", "label": "Math"}, {"key": "instructions", "label": "Instruction following"}, {"key": "code", "label": "Code"}, {"key": "german", "label": "German"}], "missing_busy": false, "missing_tps": false, "reasoning_mark": "\ud83d\udcad", "runs": [{"busy_estimated": false, "busy_h": 43.52, "label": "qwen3.8-flash-next-ple4 \u00b7 \ud83d\udcad", "run_id": 28, "scores": {"code": 78.0, "german": 80.9, "instructions": 77.3, "knowledge": 79.8, "math": 66.0, "overall": 76.4}, "tps": 20.1, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 46.07, "label": "qwen3.8-flash-next-abliterated \u00b7 \ud83d\udcad", "run_id": 34, "scores": {"code": 75.5, "german": 81.2, "instructions": 76.8, "knowledge": 82.0, "math": 65.6, "overall": 76.2}, "tps": 19.5, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 15.21, "label": "gpt-oss-120b \u00b7 \ud83d\udcad", "run_id": 18, "scores": {"code": 83.0, "german": 77.4, "instructions": 66.1, "knowledge": 68.5, "math": 70.4, "overall": 73.1}, "tps": 19.6, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 13.84, "label": "qwen3.8-flash-next-abliterated", "run_id": 35, "scores": {"code": 81.0, "german": 80.7, "instructions": 53.6, "knowledge": 77.7, "math": 71.2, "overall": 72.8}, "tps": 20.9, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 8.07, "label": "qwen3.8-flash-next-abliterated (pool)", "run_id": 32, "scores": {"code": 77.0, "german": 81.4, "instructions": 55.2, "knowledge": 77.2, "math": 72.0, "overall": 72.5}, "tps": 37.2, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 14.04, "label": "qwen3.8-flash-next-ple4", "run_id": 15, "scores": {"code": 75.5, "german": 81.1, "instructions": 49.6, "knowledge": 77.9, "math": 73.2, "overall": 71.5}, "tps": 21.6, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 12.86, "label": "gemma-4-12b-it", "run_id": 21, "scores": {"code": 75.0, "german": 74.5, "instructions": 51.3, "knowledge": 67.5, "math": 60.8, "overall": 65.8}, "tps": 13.8, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 12.53, "label": "gemma-4-12b-it-heretic", "run_id": 22, "scores": {"code": 75.0, "german": 76.5, "instructions": 50.0, "knowledge": 68.4, "math": 59.2, "overall": 65.8}, "tps": 13.8, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 20.66, "label": "lfm2.5-24b-a2b-0 \u00b7 \ud83d\udcad", "run_id": 36, "scores": {"code": 44.0, "german": 71.4, "instructions": 79.9, "knowledge": 68.1, "math": 65.2, "overall": 65.7}, "tps": 34.4, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 16.29, "label": "qwen3-coder-next", "run_id": 30, "scores": {"code": 70.5, "german": 76.0, "instructions": 45.9, "knowledge": 71.3, "math": 61.6, "overall": 65.1}, "tps": 23.8, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 12.37, "label": "qwen3.8-27b-mtp", "run_id": 20, "scores": {"code": 65.0, "german": 76.3, "instructions": 49.5, "knowledge": 69.9, "math": 60.4, "overall": 64.2}, "tps": 22.4, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 11.39, "label": "occamy-1.0", "run_id": 27, "scores": {"code": 63.0, "german": 76.9, "instructions": 44.7, "knowledge": 73.1, "math": 62.4, "overall": 64.0}, "tps": 31.6, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 12.7, "label": "qwen3.8-27b-abliterated", "run_id": 23, "scores": {"code": 59.5, "german": 72.1, "instructions": 46.1, "knowledge": 71.5, "math": 58.0, "overall": 61.4}, "tps": 21.9, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 17.46, "label": "ternary-bonsai-2-27b", "run_id": 26, "scores": {"code": 56.5, "german": 74.5, "instructions": 46.3, "knowledge": 69.5, "math": 59.6, "overall": 61.3}, "tps": 19.3, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 7.18, "label": "mimo-v2.6-flash (pool)", "run_id": 31, "scores": {"code": 61.5, "german": 74.5, "instructions": 49.8, "knowledge": 69.7, "math": 41.6, "overall": 59.4}, "tps": 21.8, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 4.09, "label": "nemotron-3.5-super-vl-120b-a12b-vllm-lilicorr", "run_id": 37, "scores": {"code": 59.5, "german": 76.9, "instructions": 53.2, "knowledge": 71.3, "math": 28.0, "overall": 57.8}, "tps": 25.3, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 13.16, "label": "gemma-4-12b-it-heretic", "run_id": 17, "scores": {"code": 62.0, "german": 67.7, "instructions": 46.8, "knowledge": 63.7, "math": 48.4, "overall": 57.7}, "tps": 21.1, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 4.41, "label": "nemotron-3.5-lightning-dspark", "run_id": 16, "scores": {"code": 53.0, "german": 70.5, "instructions": 46.0, "knowledge": 55.8, "math": 40.4, "overall": 53.1}, "tps": 70.4, "tps_source": "run"}, {"busy_estimated": true, "busy_h": 13.01, "label": "qwen3.5-9b-heretic-v2-mtp", "run_id": 19, "scores": {"code": 46.5, "german": 67.0, "instructions": 44.1, "knowledge": 65.1, "math": 35.6, "overall": 51.7}, "tps": 35.2, "tps_source": "run"}, {"busy_estimated": false, "busy_h": 9.63, "label": "zgcm-1-7b", "run_id": 29, "scores": {"code": 23.0, "german": 24.8, "instructions": 44.3, "knowledge": 34.7, "math": 40.0, "overall": 33.4}, "tps": 34.0, "tps_source": "run"}], "texts": {"busy_axis": "Compute time in hours (right = shorter)", "busy_unit": "h compute time", "estimated": "estimated", "overall": "Overall", "overall_title": "Overall score", "points": "points", "score_axis": "Score (0\u2013100)", "source_run": "measured during the run", "source_test": "from the speed test", "tps_axis": "Generation (tokens/s)", "vs": " vs. generation (tokens/s)", "vs_busy": " vs. compute time"}}</script>
 <p class="segmented" role="group" aria-label="Chart category"><button type="button" data-chart-category="overall" class="active" aria-pressed="true">Overall</button><button type="button" data-chart-category="knowledge" class="" aria-pressed="false">Knowledge &amp; reasoning</button><button type="button" data-chart-category="math" class="" aria-pressed="false">Math</button><button type="button" data-chart-category="instructions" class="" aria-pressed="false">Instruction following</button><button type="button" data-chart-category="code" class="" aria-pressed="false">Code</button><button type="button" data-chart-category="german" class="" aria-pressed="false">German</button></p>
 <div class="charts">
   <figure class="chart-card">
     <figcaption><span data-chart-title="bar">Overall score</span> <span class="sub">(💭 = with reasoning)</span></figcaption>
-    <div class="chart-box" data-chart-box="bar" style="height: 520px"><canvas data-chart="bar" role="img" aria-label="Score per model variant; the values are also in the table"></canvas></div>
+    <div class="chart-box" data-chart-box="bar" style="height: 670px"><canvas data-chart="bar" role="img" aria-label="Score per model variant; the values are also in the table"></canvas></div>
   </figure>
   <figure class="chart-card">
     <figcaption><span data-chart-title="scatter">Overall score vs. generation (tokens/s)</span> <span class="sub">(💭 = with reasoning)</span></figcaption>
-    <div class="chart-box" style="height: 320px"><canvas data-chart="scatter" role="img" aria-label="Score against generation speed in tokens per second; the values are also in the table"></canvas></div>
-    <p class="hint">x-axis: generation per request measured during the run; hollow points only have the speed-test value.</p>
+    <div class="segmented" role="group" aria-label="Scatter chart x-axis"><button type="button" data-chart-x="tps" class="active" aria-pressed="true">Tokens/s</button><button type="button" data-chart-x="busy" aria-pressed="false">Compute time</button></div>
+    <div class="chart-box" style="height: 320px"><canvas data-chart="scatter" role="img" aria-label="Score against generation speed in tokens per second or against compute time; the values are also in the table"></canvas></div>
+    <p class="hint" data-chart-hint="tps">x-axis: generation per request measured during the run; hollow points only have the speed-test value.</p>
+    <p class="hint" data-chart-hint="busy" hidden>x-axis: compute time for the whole suite, mirrored – the further right, the sooner the model is done with the tasks. Hollow points are estimated (tasks without a request log count with their wall-clock duration). Comparable on the same hardware.</p>
   </figure>
 </div>
 </div>
 </section>
-<p class="hint">Scores run from 0 to 100; the best value per category is shown in bold. Hovering a category cell lists its tasks with standard error (±) and number of items (n). Speed: generation in tokens/s from the speed test and measured during the run (gen t/s run, per request), prompt processing at 4k context in tokens/s, time to first token at 4k context in ms. Click a column header to sort.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<p class="hint">Scores run from 0 to 100; the best value per category is shown in bold. Hovering a category cell lists its tasks with standard error (±) and number of items (n). Speed: generation in tokens/s from the speed test and measured during the run (gen t/s run, per request), prompt processing at 4k context in tokens/s, time to first token at 4k context in ms. Compute time: hours during which the run's requests were in flight, excluding pauses and reruns (≈ = estimated). Click a column header to sort.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="chart.umd.min.js"></script>
 <script src="app.js"></script>

methodology.html

@@ -35,6 +35,11 @@ context lengths, each as the median of several repetitions with p10 and p90. Dep
 is taken directly at the inference server or through the AI hub; values measured through the AI hub include queue
 and proxy time and are therefore somewhat slower. The route is stated on the run's own page. Speed depends heavily
 on hardware and inference software.</p>
+<p>Tokens per second do not tell the whole story: a model that writes or thinks a lot is slow to finish despite a
+high rate. <strong>Compute time</strong> measures this directly – the time during which the run's requests were in
+flight, summed over all tasks of the suite, with parallel requests counted once and pauses and reruns left out. It
+depends on the hardware and on the number of parallel requests. Where a task has no request log (older runs), its
+wall-clock duration stands in and the value is marked as estimated (≈).</p>
 </section>
 
 <h2>Suite v2b</h2>
@@ -117,7 +122,7 @@ on hardware and inference software.</p>
 </tbody>
 </table>
 </div>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="chart.umd.min.js"></script>
 <script src="app.js"></script>

runs/15.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">71.5</dd>
+  <dt>Compute time</dt><dd>≈ 14.0 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/16.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">53.1</dd>
+  <dt>Compute time</dt><dd>≈ 4.4 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/17.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">57.7</dd>
+  <dt>Compute time</dt><dd>≈ 13.2 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/18.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">73.1</dd>
+  <dt>Compute time</dt><dd>15.2 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -126,7 +127,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/19.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">51.7</dd>
+  <dt>Compute time</dt><dd>≈ 13.0 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/20.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">64.2</dd>
+  <dt>Compute time</dt><dd>≈ 12.4 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/21.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">65.8</dd>
+  <dt>Compute time</dt><dd>≈ 12.9 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/22.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">65.8</dd>
+  <dt>Compute time</dt><dd>≈ 12.5 h <span class="sub">(estimated: tasks without a request log count with their wall-clock duration)</span></dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/23.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">61.4</dd>
+  <dt>Compute time</dt><dd>12.7 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/26.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">61.3</dd>
+  <dt>Compute time</dt><dd>17.5 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/27.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">64.0</dd>
+  <dt>Compute time</dt><dd>11.4 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/28.html

@@ -0,0 +1,197 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.8-flash-next-ple4 – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>qwen3.8-flash-next-ple4</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>W4A16</dd>
+  <dt>Reasoning</dt><dd>on</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-09-21 – 2026-10-12</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">76.4</dd>
+  <dt>Compute time</dt><dd>43.5 h</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">79.8</td></tr>
+<tr><td>Math</td><td class="num">66.0</td></tr>
+<tr><td>Instruction following</td><td class="num">77.3</td></tr>
+<tr><td>Code</td><td class="num">78.0</td></tr>
+<tr><td>German</td><td class="num">80.9</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">89.3</td>
+  <td class="num">± 1.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">75.8</td>
+  <td class="num">± 3.1</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">74.4</td>
+  <td class="num">± 2.8</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">66.0</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">71.3</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">95.3</td>
+  <td class="num">± 1.7</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">78.0</td>
+  <td class="num">± 2.9</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">86.0</td>
+  <td class="num">± 2.0</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">70.8</td>
+  <td class="num">± 3.7</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">25.2</td>
+  <td class="num">24.6</td>
+  <td class="num">25.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">552</td>
+  <td class="num">1032.8</td>
+  <td class="num">731.8</td>
+  <td class="num">1038.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">3908</td>
+  <td class="num">1952.3</td>
+  <td class="num">1904.8</td>
+  <td class="num">1963.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15426</td>
+  <td class="num">2152.8</td>
+  <td class="num">2068.1</td>
+  <td class="num">2153.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">552</td>
+  <td class="num">534.5</td>
+  <td class="num">531.8</td>
+  <td class="num">807.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">3908</td>
+  <td class="num">2001.7</td>
+  <td class="num">1990.5</td>
+  <td class="num">2052.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15426</td>
+  <td class="num">7165.6</td>
+  <td class="num">7161.8</td>
+  <td class="num">7466.6</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../chart.umd.min.js"></script>
+<script src="../app.js"></script>
+</body>
+</html>

runs/29.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">33.4</dd>
+  <dt>Compute time</dt><dd>9.6 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/30.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">65.1</dd>
+  <dt>Compute time</dt><dd>16.3 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/31.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">59.4</dd>
+  <dt>Compute time</dt><dd>7.2 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/32.html

@@ -27,6 +27,7 @@
   <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
   <dt>Measured</dt><dd>directly at the inference server</dd>
   <dt>Overall</dt><dd class="strong">72.5</dd>
+  <dt>Compute time</dt><dd>8.1 h</dd>
 </dl>
 
 <h2>Categories</h2>
@@ -188,7 +189,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-10-10 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../chart.umd.min.js"></script>
 <script src="../app.js"></script>

runs/34.html

@@ -0,0 +1,197 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.8-flash-next-abliterated – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>qwen3.8-flash-next-abliterated</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>iQ3-S</dd>
+  <dt>Reasoning</dt><dd>on</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-10-11 – 2026-10-12</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">76.2</dd>
+  <dt>Compute time</dt><dd>46.1 h</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">82.0</td></tr>
+<tr><td>Math</td><td class="num">65.6</td></tr>
+<tr><td>Instruction following</td><td class="num">76.8</td></tr>
+<tr><td>Code</td><td class="num">75.5</td></tr>
+<tr><td>German</td><td class="num">81.2</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">89.0</td>
+  <td class="num">± 1.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">74.7</td>
+  <td class="num">± 3.1</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">82.4</td>
+  <td class="num">± 2.4</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">65.6</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">71.3</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">93.3</td>
+  <td class="num">± 2.0</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">75.5</td>
+  <td class="num">± 3.0</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">86.7</td>
+  <td class="num">± 2.0</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">70.1</td>
+  <td class="num">± 3.7</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">34.8</td>
+  <td class="num">23.4</td>
+  <td class="num">37.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">552</td>
+  <td class="num">355.8</td>
+  <td class="num">354.5</td>
+  <td class="num">357.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">3908</td>
+  <td class="num">575.3</td>
+  <td class="num">569.6</td>
+  <td class="num">577.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15426</td>
+  <td class="num">604.6</td>
+  <td class="num">598.9</td>
+  <td class="num">624.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">552</td>
+  <td class="num">1551.5</td>
+  <td class="num">1544.2</td>
+  <td class="num">1557.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">3908</td>
+  <td class="num">6792.9</td>
+  <td class="num">6762.9</td>
+  <td class="num">6860.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15426</td>
+  <td class="num">25513.6</td>
+  <td class="num">24709.1</td>
+  <td class="num">25756.6</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../chart.umd.min.js"></script>
+<script src="../app.js"></script>
+</body>
+</html>

runs/35.html

@@ -0,0 +1,197 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.8-flash-next-abliterated – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>qwen3.8-flash-next-abliterated</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>iQ3-S</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-10-12 – 2026-10-13</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">72.8</dd>
+  <dt>Compute time</dt><dd>13.8 h</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">77.7</td></tr>
+<tr><td>Math</td><td class="num">71.2</td></tr>
+<tr><td>Instruction following</td><td class="num">53.6</td></tr>
+<tr><td>Code</td><td class="num">81.0</td></tr>
+<tr><td>German</td><td class="num">80.7</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">83.7</td>
+  <td class="num">± 2.1</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">77.8</td>
+  <td class="num">± 3.0</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">71.6</td>
+  <td class="num">± 2.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">71.2</td>
+  <td class="num">± 2.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">42.0</td>
+  <td class="num">± 2.9</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">88.7</td>
+  <td class="num">± 2.6</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">81.0</td>
+  <td class="num">± 2.8</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">84.7</td>
+  <td class="num">± 2.1</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">72.7</td>
+  <td class="num">± 3.6</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">38.1</td>
+  <td class="num">35.6</td>
+  <td class="num">38.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">519.4</td>
+  <td class="num">513.2</td>
+  <td class="num">523.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">658.3</td>
+  <td class="num">657.7</td>
+  <td class="num">659.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">644.3</td>
+  <td class="num">643.2</td>
+  <td class="num">652.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">1018.6</td>
+  <td class="num">1010.6</td>
+  <td class="num">1029.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">6113.0</td>
+  <td class="num">6101.9</td>
+  <td class="num">6118.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">24835.5</td>
+  <td class="num">24508.9</td>
+  <td class="num">24876.7</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../chart.umd.min.js"></script>
+<script src="../app.js"></script>
+</body>
+</html>

runs/36.html

@@ -0,0 +1,197 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>lfm2.5-24b-a2b-0 – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>lfm2.5-24b-a2b-0</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>Q8</dd>
+  <dt>Reasoning</dt><dd>on</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-10-20 – 2026-10-21</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>llama.cpp b10014</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">65.7</dd>
+  <dt>Compute time</dt><dd>20.7 h</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">68.1</td></tr>
+<tr><td>Math</td><td class="num">65.2</td></tr>
+<tr><td>Instruction following</td><td class="num">79.9</td></tr>
+<tr><td>Code</td><td class="num">44.0</td></tr>
+<tr><td>German</td><td class="num">71.4</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">76.7</td>
+  <td class="num">± 2.4</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">59.1</td>
+  <td class="num">± 3.5</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">68.4</td>
+  <td class="num">± 2.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">65.2</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">75.0</td>
+  <td class="num">± 2.5</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">94.7</td>
+  <td class="num">± 1.8</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">44.0</td>
+  <td class="num">± 3.5</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">75.0</td>
+  <td class="num">± 2.5</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">64.3</td>
+  <td class="num">± 3.9</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">77.1</td>
+  <td class="num">77.0</td>
+  <td class="num">77.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">525</td>
+  <td class="num">2017.0</td>
+  <td class="num">1952.9</td>
+  <td class="num">2405.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4027</td>
+  <td class="num">2775.2</td>
+  <td class="num">2772.5</td>
+  <td class="num">2807.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16044</td>
+  <td class="num">2732.4</td>
+  <td class="num">2719.1</td>
+  <td class="num">2756.8</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">525</td>
+  <td class="num">260.3</td>
+  <td class="num">218.3</td>
+  <td class="num">268.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4027</td>
+  <td class="num">1451.1</td>
+  <td class="num">1434.5</td>
+  <td class="num">1452.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16044</td>
+  <td class="num">5871.9</td>
+  <td class="num">5820.0</td>
+  <td class="num">5900.7</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../chart.umd.min.js"></script>
+<script src="../app.js"></script>
+</body>
+</html>

runs/37.html

@@ -0,0 +1,197 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>nemotron-3.5-super-vl-120b-a12b-vllm-lilicorr – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>nemotron-3.5-super-vl-120b-a12b-vllm-lilicorr</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-10-21 – 2026-10-22</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">57.8</dd>
+  <dt>Compute time</dt><dd>4.1 h</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">71.3</td></tr>
+<tr><td>Math</td><td class="num">28.0</td></tr>
+<tr><td>Instruction following</td><td class="num">53.2</td></tr>
+<tr><td>Code</td><td class="num">59.5</td></tr>
+<tr><td>German</td><td class="num">76.9</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">81.7</td>
+  <td class="num">± 2.2</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">68.7</td>
+  <td class="num">± 3.3</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">63.6</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">28.0</td>
+  <td class="num">± 2.8</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">40.7</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">91.3</td>
+  <td class="num">± 2.3</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">59.5</td>
+  <td class="num">± 3.5</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">79.7</td>
+  <td class="num">± 2.3</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">71.4</td>
+  <td class="num">± 3.7</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">27.5</td>
+  <td class="num">26.1</td>
+  <td class="num">28.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">1241.7</td>
+  <td class="num">1230.8</td>
+  <td class="num">1249.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4008</td>
+  <td class="num">2878.4</td>
+  <td class="num">2869.1</td>
+  <td class="num">2890.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15949</td>
+  <td class="num">3134.4</td>
+  <td class="num">2901.2</td>
+  <td class="num">3153.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">426.0</td>
+  <td class="num">423.2</td>
+  <td class="num">429.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4008</td>
+  <td class="num">1392.5</td>
+  <td class="num">1386.6</td>
+  <td class="num">1396.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15949</td>
+  <td class="num">5088.5</td>
+  <td class="num">5058.0</td>
+  <td class="num">5518.0</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-10-22 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../chart.umd.min.js"></script>
+<script src="../app.js"></script>
+</body>
+</html>

style.css

@@ -111,6 +111,7 @@ th[aria-sort="descending"]::after { content: " ▼"; }
 .charts { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(100%, 480px), 1fr)); gap: 16px; margin-top: 16px; }
 .chart-card { margin: 0 0 16px; padding: 12px; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
 figcaption { font-weight: 600; margin-bottom: 8px; }
+.chart-card .segmented { display: flex; flex-wrap: wrap; margin: 0 0 8px; font-size: 12px; }
 .chart-box { position: relative; width: 100%; }
 .thresholds { display: flex; flex-wrap: wrap; gap: 8px 12px; align-items: center; margin: 0; }
 .thresholds label { display: inline-flex; gap: 6px; align-items: center; color: var(--ink-2); }