Ergebnisse aktualisiert (5 Läufe)

README.md

@@ -1,11 +1,14 @@
 ---
-title: LocalLLMBenchmarks
-emoji: 🐨
-colorFrom: pink
-colorTo: red
+title: Local LLM Benchmarks
+emoji: 📊
+colorFrom: blue
+colorTo: gray
 sdk: static
+app_file: index.html
 pinned: false
-short_description: Benchmarks of local Modells (running at NVIDIA Spark)
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Local LLM Benchmarks
+
+Ergebnisse lokaler Sprachmodelle (Wissen, Mathe, Code, Deutsch) als statische Seite. Die Seite wird mit
+`bench publish` erzeugt; die Rohdaten stehen in `results.json` (Schema-Version 1).

app.js

@@ -0,0 +1,198 @@
+(function () {
+  "use strict";
+
+  // Tabellen sortieren (Klick auf th[data-sort]); leere Werte stehen immer am Ende.
+  document.querySelectorAll("table.sortable").forEach(function (table) {
+    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); });
+      });
+    });
+  });
+
+  // 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;
+    });
+  });
+
+  // Bestätigung vor gefährlichen Aktionen (Formulare mit data-confirm).
+  document.querySelectorAll("form[data-confirm]").forEach(function (form) {
+    form.addEventListener("submit", function (event) {
+      if (!window.confirm(form.dataset.confirm)) event.preventDefault();
+    });
+  });
+
+  if (typeof Chart === "undefined") return;
+
+  var css = getComputedStyle(document.documentElement);
+  function token(name) { return css.getPropertyValue(name).trim(); }
+  var ink = token("--ink"), ink2 = token("--ink-2"), muted = token("--muted");
+  var grid = token("--grid"), axisColor = token("--axis"), surface = token("--surface");
+  var series = [token("--series-1"), token("--series-2"), token("--series-3")];
+  var fontFamily = 'system-ui, -apple-system, "Segoe UI", sans-serif';
+  Chart.defaults.font.family = fontFamily;
+  Chart.defaults.color = ink2;
+
+  function axis(title, extra) {
+    var config = {
+      grid: { color: grid, drawTicks: false },
+      border: { color: axisColor },
+      ticks: { color: muted, padding: 6 }
+    };
+    if (title) config.title = { display: true, text: title, color: ink2 };
+    return Object.assign(config, extra || {});
+  }
+  function readJson(id) {
+    var el = document.getElementById(id);
+    return el ? JSON.parse(el.textContent) : null;
+  }
+
+  // Wert direkt am Balkenende (nur für die einreihige Übersicht).
+  var barValueLabels = {
+    id: "barValueLabels",
+    afterDatasetsDraw: function (chart) {
+      var meta = chart.getDatasetMeta(0);
+      var ctx = chart.ctx;
+      ctx.save();
+      ctx.fillStyle = ink;
+      ctx.font = "12px " + fontFamily;
+      ctx.textBaseline = "middle";
+      meta.data.forEach(function (bar, i) {
+        var value = chart.data.datasets[0].data[i];
+        if (value !== null && value !== undefined) ctx.fillText(value.toFixed(1), bar.x + 6, bar.y);
+      });
+      ctx.restore();
+    }
+  };
+
+  var overview = readJson("overview-data");
+  if (overview && overview.bars.length) {
+    new Chart(document.getElementById("bar-chart"), {
+      type: "bar",
+      data: {
+        labels: overview.bars.map(function (b) { return b.label; }),
+        datasets: [{
+          label: "Gesamtscore",
+          data: overview.bars.map(function (b) { return b.value; }),
+          backgroundColor: series[0],
+          borderRadius: 4,
+          borderSkipped: "start",
+          barPercentage: 0.8,
+          categoryPercentage: 0.9,
+          maxBarThickness: 18
+        }]
+      },
+      options: {
+        indexAxis: "y",
+        maintainAspectRatio: false,
+        animation: false,
+        layout: { padding: { right: 40 } },
+        scales: { x: axis("Score (0–100)", { min: 0, max: 100 }), y: axis(null, { grid: { display: false } }) },
+        plugins: {
+          legend: { display: false },
+          tooltip: { callbacks: { label: function (c) { return "Gesamt: " + c.parsed.x.toFixed(1); } } }
+        }
+      },
+      plugins: [barValueLabels]
+    });
+  }
+
+  if (overview && overview.scatter.length) {
+    new Chart(document.getElementById("scatter-chart"), {
+      type: "scatter",
+      data: {
+        datasets: [{
+          label: "Modellvarianten",
+          data: overview.scatter,
+          backgroundColor: series[0],
+          borderColor: surface,
+          borderWidth: 2,
+          pointRadius: 5,
+          pointHoverRadius: 7,
+          pointHitRadius: 12
+        }]
+      },
+      options: {
+        maintainAspectRatio: false,
+        animation: false,
+        scales: { x: axis("Generierung (Tokens/s)", { beginAtZero: true }), y: axis("Gesamtscore", { min: 0, max: 100 }) },
+        plugins: {
+          legend: { display: false },
+          tooltip: {
+            callbacks: {
+              label: function (c) { return c.raw.label + ": " + c.raw.y.toFixed(1) + " Punkte, " + c.raw.x.toFixed(1) + " t/s"; }
+            }
+          }
+        }
+      }
+    });
+  }
+
+  var categories = readJson("category-data");
+  if (categories && categories.series.length) {
+    new Chart(document.getElementById("category-chart"), {
+      type: "bar",
+      data: {
+        labels: categories.labels,
+        datasets: categories.series.map(function (s, i) {
+          return {
+            label: s.label,
+            data: s.values,
+            backgroundColor: series[i],
+            borderColor: surface,
+            borderWidth: { top: 1, bottom: 1, right: 2 },
+            borderRadius: 4,
+            borderSkipped: "start",
+            barPercentage: 0.9,
+            categoryPercentage: 0.8,
+            maxBarThickness: 16
+          };
+        })
+      },
+      options: {
+        indexAxis: "y",
+        maintainAspectRatio: false,
+        animation: false,
+        scales: { x: axis("Score (0–100)", { min: 0, max: 100 }), y: axis(null, { grid: { display: false } }) },
+        plugins: {
+          legend: { display: true, position: "top", align: "start", labels: { usePointStyle: true, pointStyle: "rectRounded", color: ink2 } },
+          tooltip: {
+            callbacks: {
+              label: function (c) { return c.dataset.label + ": " + (c.parsed.x === null ? "–" : c.parsed.x.toFixed(1)); }
+            }
+          }
+        }
+      }
+    });
+  }
+})();

index.html

@@ -1,19 +1,115 @@
 <!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>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" class="active">Übersicht</a>
+    <a href="methodik.html">Methodik</a>
+    <a href="results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<h1>Übersicht</h1>
+<p class="lead">Lokale Sprachmodelle im Vergleich: Wissen, Mathe, Code und Deutsch, dazu die Geschwindigkeit. Je Modellvariante steht der neueste fertige Lauf. Wie gemessen wird, steht unter <a href="methodik.html">Methodik</a>.</p>
+<section id="v2-quick">
+<h2>Suite v2 · Profil quick</h2>
+<div class="table-wrap">
+<table class="sortable">
+<thead><tr>
+  <th data-sort="text">Modell</th>
+  <th data-sort="num" class="num">Gesamt</th>
+<th data-sort="num" class="num">Wissen &amp; Reasoning</th><th data-sort="num" class="num">Mathe &amp; Instruktionen</th><th data-sort="num" class="num">Code</th><th data-sort="num" class="num">Deutsch</th><th data-sort="num" class="num">gen t/s</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">Datum</th>
+</tr></thead>
+<tbody>
+<tr>
+  <td data-value="qwen3.8-flash-next-ple4">
+    <a href="runs/8.html" class="strong">qwen3.8-flash-next-ple4</a>
+    <div class="sub">NVFP4 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="87.66702376525996">87.7</td>
+  <td class="num strong" data-value="81.8030303030303" title="mmlu_pro: 86.3 ± 2.0 (n=300)&#10;gpqa_diamond_gen: 77.3 ± 3.0 (n=198)&#10;">81.8</td>
+  <td class="num strong" data-value="95.28888888888889" title="math_hard_gen: 97.6 ± 1.0 (n=250)&#10;ifeval: 90.7 ± 2.4 (n=150)&#10;">95.3</td>
+  <td class="num strong" data-value="87.90950920245399" title="humaneval_plus_gen: 96.3 ± 1.5 (n=163)&#10;mbpp_plus_gen: 79.5 ± 2.9 (n=200)&#10;">87.9</td>
+  <td class="num strong" data-value="85.66666666666667" title="mmlu_prox_de: 85.7 ± 2.0 (n=300)&#10;">85.7</td>
+  <td class="num" data-value="26.376464212349354">26</td>
+  <td class="num" data-value="2041.4640201348918">2041</td>
+  <td class="num" data-value="1973.583644023165">1974</td>
+  <td data-value="2026-09-15">2026-09-15</td>
+</tr>
+<tr>
+  <td data-value="qwen3.8-27b-mtp">
+    <a href="runs/10.html" class="strong">qwen3.8-27b-mtp</a>
+    <div class="sub">NVFP4 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="80.69203693375472">80.7</td>
+  <td class="num" data-value="71.98484848484848" title="mmlu_pro: 80.3 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 63.6 ± 3.4 (n=198)&#10;">72.0</td>
+  <td class="num" data-value="88.57777777777778" title="math_hard_gen: 91.2 ± 1.8 (n=250)&#10;ifeval: 83.3 ± 3.1 (n=150)&#10;">88.6</td>
+  <td class="num" data-value="83.20552147239263" title="humaneval_plus_gen: 91.4 ± 2.2 (n=163)&#10;mbpp_plus_gen: 75.0 ± 3.1 (n=200)&#10;">83.2</td>
+  <td class="num" data-value="79.0" title="mmlu_prox_de: 79.0 ± 2.4 (n=300)&#10;">79.0</td>
+  <td class="num" data-value="15.103805505442727">15</td>
+  <td class="num" data-value="2550.5619640606633">2551</td>
+  <td class="num" data-value="1577.6915270835161">1578</td>
+  <td data-value="2026-09-16">2026-09-16</td>
+</tr>
+<tr>
+  <td data-value="gemma-4-12b-it-heretic">
+    <a href="runs/7.html" class="strong">gemma-4-12b-it-heretic</a>
+    <div class="sub">NVFP4 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="76.9312805973849">76.9</td>
+  <td class="num" data-value="66.36363636363636" title="mmlu_pro: 76.7 ± 2.4 (n=300)&#10;gpqa_diamond_gen: 56.1 ± 3.5 (n=198)&#10;">66.4</td>
+  <td class="num" data-value="86.62222222222222" title="math_hard_gen: 85.6 ± 2.2 (n=250)&#10;ifeval: 88.7 ± 2.6 (n=150)&#10;">86.6</td>
+  <td class="num" data-value="84.73926380368098" title="humaneval_plus_gen: 94.5 ± 1.8 (n=163)&#10;mbpp_plus_gen: 75.0 ± 3.1 (n=200)&#10;">84.7</td>
+  <td class="num" data-value="70.0" title="mmlu_prox_de: 70.0 ± 2.7 (n=300)&#10;">70.0</td>
+  <td class="num" data-value="21.360681311630678">21</td>
+  <td class="num" data-value="5258.027383235864">5258</td>
+  <td class="num" data-value="763.404164230451">763</td>
+  <td data-value="2026-09-15">2026-09-15</td>
+</tr>
+<tr>
+  <td data-value="nemotron-3.5-lightning-dspark">
+    <a href="runs/6.html" class="strong">nemotron-3.5-lightning-dspark</a>
+    <div class="sub">NVFP4 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="76.34037383032782">76.3</td>
+  <td class="num" data-value="64.03535353535354" title="mmlu_pro: 71.0 ± 2.6 (n=300)&#10;gpqa_diamond_gen: 57.1 ± 3.5 (n=198)&#10;">64.0</td>
+  <td class="num" data-value="86.71111111111111" title="math_hard_gen: 90.4 ± 1.9 (n=250)&#10;ifeval: 79.3 ± 3.3 (n=150)&#10;">86.7</td>
+  <td class="num" data-value="82.61503067484662" title="humaneval_plus_gen: 87.7 ± 2.6 (n=163)&#10;mbpp_plus_gen: 77.5 ± 3.0 (n=200)&#10;">82.6</td>
+  <td class="num" data-value="72.0" title="mmlu_prox_de: 72.0 ± 2.6 (n=300)&#10;">72.0</td>
+  <td class="num" data-value="83.7975869946743">84</td>
+  <td class="num" data-value="7236.849600555504">7237</td>
+  <td class="num" data-value="554.1085170116276">554</td>
+  <td data-value="2026-09-15">2026-09-15</td>
+</tr>
+<tr>
+  <td data-value="qwen3.5-9b-heretic-v2-mtp">
+    <a href="runs/11.html" class="strong">qwen3.5-9b-heretic-v2-mtp</a>
+    <div class="sub">NVFP4 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="73.63538219619508">73.6</td>
+  <td class="num" data-value="67.2070707070707" title="mmlu_pro: 76.3 ± 2.5 (n=300)&#10;gpqa_diamond_gen: 58.1 ± 3.5 (n=198)&#10;">67.2</td>
+  <td class="num" data-value="80.53333333333333" title="math_hard_gen: 76.8 ± 2.7 (n=250)&#10;ifeval: 88.0 ± 2.7 (n=150)&#10;">80.5</td>
+  <td class="num" data-value="77.46779141104295" title="humaneval_plus_gen: 83.4 ± 2.9 (n=163)&#10;mbpp_plus_gen: 71.5 ± 3.2 (n=200)&#10;">77.5</td>
+  <td class="num" data-value="69.33333333333334" title="mmlu_prox_de: 69.3 ± 2.7 (n=300)&#10;">69.3</td>
+  <td class="num" data-value="31.530246994028012">32</td>
+  <td class="num" data-value="6290.925546819461">6291</td>
+  <td class="num" data-value="639.6515059750527">640</td>
+  <td data-value="2026-09-16">2026-09-16</td>
+</tr>
+</tbody>
+</table>
+</div>
+</section>
+<p class="hint">Scores 0–100; fett ist der Bestwert je Kategorie. Der Mauszeiger über einer Kategorie zeigt die Tasks mit Standardfehler (±) und Anzahl Aufgaben (n). Geschwindigkeit: Generierung in Tokens/s, Prompt-Verarbeitung bei 4k Kontext in Tokens/s, Zeit bis zum ersten Token bei 4k Kontext in ms. Klick auf eine Spaltenüberschrift sortiert.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="app.js"></script>
+</body>
 </html>

methodik.html

@@ -0,0 +1,109 @@
+<!doctype html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Methodik – 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">Übersicht</a>
+    <a href="methodik.html" class="active">Methodik</a>
+    <a href="results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<h1>Methodik</h1>
+<section class="prose">
+<p>Jedes Modell bearbeitet dieselben öffentlichen Aufgabensammlungen (Tasks). Ein Task liefert einen Score von 0 bis 100.
+Der Score einer Kategorie ist das gewichtete Mittel ihrer Tasks, der Gesamtscore das ungewichtete Mittel der Kategorien.
+Fehlt ein Task, bleiben Kategorie und Gesamt leer.</p>
+<ul>
+  <li><strong>Profil quick:</strong> feste Stichprobe je Task (Spalte quick_n), für alle Modelle dieselben Aufgaben.</li>
+  <li><strong>Profil full:</strong> alle Aufgaben des Tasks.</li>
+  <li><strong>Suite-Version:</strong> ändern sich Tasks, Prompts oder Metriken, entsteht eine neue Version. Scores verschiedener Versionen sind nicht vergleichbar.</li>
+</ul>
+<h2>Unsicherheit</h2>
+<p>Zu jedem Task-Score gehört ein Standardfehler (±). Das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.
+Überlappen sich die Intervalle zweier Modelle, ist der Unterschied nicht belastbar – gerade im Profil quick mit
+kleineren Stichproben.</p>
+<h2>Geschwindigkeit</h2>
+<p>Gemessen werden Generierung (Tokens/s), Prompt-Verarbeitung (Tokens/s) und die Zeit bis zum ersten Token bei
+verschiedenen Kontextlängen, jeweils als Median mehrerer Wiederholungen mit p10 und p90. Je nach Lauf wird direkt am
+Inferenz-Server oder über den AI-Hub gemessen; über den AI-Hub gemessene Werte enthalten Queue- und Proxyzeit und
+sind daher etwas langsamer. Der Messweg steht auf der Seite des Laufs. Geschwindigkeit hängt stark von Hardware und
+Inferenz-Software ab.</p>
+</section>
+
+<h2>Suite v2</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th>Metrik</th><th>Art</th><th class="num">Gewicht</th><th class="num">quick_n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>mmlu_pro</td>
+  <td>exact_match,custom-extract</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td>exact_match,letter</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td>exact_match,boxed</td>
+  <td>Text</td>
+  <td class="num">2</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>ifeval</td>
+  <td>prompt_level_strict_acc,none</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>humaneval_plus_gen</td>
+  <td>pass@1</td>
+  <td>Code (ausgeführt in Sandbox)</td>
+  <td class="num">1</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td>pass@1</td>
+  <td>Code (ausgeführt in Sandbox)</td>
+  <td class="num">1</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td>exact_match,custom-extract</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="app.js"></script>
+</body>
+</html>

results.json

@@ -0,0 +1,866 @@
+{
+  "schema_version": 1,
+  "generator": "bench publish",
+  "generated_at": "2026-09-17",
+  "suites": [
+    {
+      "version": 2,
+      "label": "v2",
+      "categories": [
+        {
+          "key": "knowledge",
+          "label": "Wissen & Reasoning",
+          "tasks": [
+            {
+              "name": "mmlu_pro",
+              "metric": "exact_match,custom-extract",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 300
+            },
+            {
+              "name": "gpqa_diamond_gen",
+              "metric": "exact_match,letter",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 198
+            }
+          ]
+        },
+        {
+          "key": "math",
+          "label": "Mathe & Instruktionen",
+          "tasks": [
+            {
+              "name": "math_hard_gen",
+              "metric": "exact_match,boxed",
+              "kind": "lm",
+              "weight": 2.0,
+              "quick_n": 250
+            },
+            {
+              "name": "ifeval",
+              "metric": "prompt_level_strict_acc,none",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 150
+            }
+          ]
+        },
+        {
+          "key": "code",
+          "label": "Code",
+          "tasks": [
+            {
+              "name": "humaneval_plus_gen",
+              "metric": "pass@1",
+              "kind": "code",
+              "weight": 1.0,
+              "quick_n": 163
+            },
+            {
+              "name": "mbpp_plus_gen",
+              "metric": "pass@1",
+              "kind": "code",
+              "weight": 1.0,
+              "quick_n": 200
+            }
+          ]
+        },
+        {
+          "key": "german",
+          "label": "Deutsch",
+          "tasks": [
+            {
+              "name": "mmlu_prox_de",
+              "metric": "exact_match,custom-extract",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 300
+            }
+          ]
+        }
+      ]
+    }
+  ],
+  "overview": [
+    {
+      "suite_version": 2,
+      "profile": "quick",
+      "runs": [
+        8,
+        10,
+        7,
+        6,
+        11
+      ]
+    }
+  ],
+  "runs": [
+    {
+      "id": 6,
+      "model": "nemotron-3.5-lightning-dspark",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-15",
+      "finished_at": "2026-09-15",
+      "adopted_from_date": null,
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 76.34037383032782,
+      "categories": {
+        "knowledge": 64.03535353535354,
+        "math": 86.71111111111111,
+        "code": 82.61503067484662,
+        "german": 72.0
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 71.0,
+          "stderr": 2.624173703362272,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 57.07070707070707,
+          "stderr": 3.5265527246012,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 90.4,
+          "stderr": 1.8668961419477186,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 79.33333333333333,
+          "stderr": 3.3171868444173502,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 87.73006134969326,
+          "stderr": 2.5777328426978916,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 77.5,
+          "stderr": 2.9601626330440616,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 72.0,
+          "stderr": 2.5966276044877823,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 83.7975869946743,
+          "p10": 80.73469344921789,
+          "p90": 87.48503158643007,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 528,
+          "value": 3738.2272618348616,
+          "p10": 2851.3085621807413,
+          "p90": 3772.3639705377145,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4010,
+          "value": 7236.849600555504,
+          "p10": 6862.405850690909,
+          "p90": 7270.225969572038,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15960,
+          "value": 7177.337804243572,
+          "p10": 7129.311202773395,
+          "p90": 7180.000227156906,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 528,
+          "value": 141.2434191443026,
+          "p10": 139.96619619429111,
+          "p90": 185.03863401710987,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4010,
+          "value": 554.1085170116276,
+          "p10": 551.5700471587479,
+          "p90": 585.4010013863444,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15960,
+          "value": 2223.665714962408,
+          "p10": 2222.8412566240877,
+          "p90": 2238.6918489821255,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 7,
+      "model": "gemma-4-12b-it-heretic",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-15",
+      "finished_at": "2026-09-15",
+      "adopted_from_date": null,
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 76.9312805973849,
+      "categories": {
+        "knowledge": 66.36363636363636,
+        "math": 86.62222222222222,
+        "code": 84.73926380368098,
+        "german": 70.0
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 76.66666666666667,
+          "stderr": 2.445997952351143,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 56.060606060606055,
+          "stderr": 3.5360859475294797,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 85.6,
+          "stderr": 2.2249407735450246,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 88.66666666666667,
+          "stderr": 2.5969637523072806,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 94.47852760736197,
+          "stderr": 1.7944712448654647,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 75.0,
+          "stderr": 3.0695456590127175,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 70.0,
+          "stderr": 2.6501719512585327,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 21.360681311630678,
+          "p10": 21.351851046943292,
+          "p90": 21.37119739804449,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 528,
+          "value": 5070.531579450607,
+          "p10": 4947.684367849902,
+          "p90": 5118.2855736092215,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4014,
+          "value": 5258.027383235864,
+          "p10": 5221.741910086158,
+          "p90": 5293.868867392755,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15985,
+          "value": 3514.090286443633,
+          "p10": 3505.209867577695,
+          "p90": 3518.510081028629,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 528,
+          "value": 103.93387591466308,
+          "p10": 103.16019831225276,
+          "p90": 106.71658930368721,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4014,
+          "value": 763.404164230451,
+          "p10": 758.2495824899524,
+          "p90": 768.7096168287098,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15985,
+          "value": 4548.83019416593,
+          "p10": 4543.119800603017,
+          "p90": 4560.354916239157,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 8,
+      "model": "qwen3.8-flash-next-ple4",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-15",
+      "finished_at": "2026-09-16",
+      "adopted_from_date": null,
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 87.66702376525996,
+      "categories": {
+        "knowledge": 81.8030303030303,
+        "math": 95.28888888888889,
+        "code": 87.90950920245399,
+        "german": 85.66666666666667
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 86.33333333333333,
+          "stderr": 1.986483497350814,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 77.27272727272727,
+          "stderr": 2.9857515673386437,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 97.6,
+          "stderr": 0.9699087026964269,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 90.66666666666666,
+          "stderr": 2.383137380283291,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 96.31901840490798,
+          "stderr": 1.4793820323252016,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 79.5,
+          "stderr": 2.861764926136022,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 85.66666666666667,
+          "stderr": 2.0264875529605346,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 26.376464212349354,
+          "p10": 26.252651590876688,
+          "p90": 26.55509547793904,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 1051.1180216767418,
+          "p10": 1048.4388391083346,
+          "p90": 1056.205300748408,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4029,
+          "value": 2041.4640201348918,
+          "p10": 2032.85099803181,
+          "p90": 2044.4743063547462,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 16012,
+          "value": 2266.690263749667,
+          "p10": 2263.078563697444,
+          "p90": 2267.8667843911776,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 503.273646812886,
+          "p10": 500.2890304662287,
+          "p90": 504.55993725918233,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4029,
+          "value": 1973.583644023165,
+          "p10": 1970.680379960686,
+          "p90": 1981.9602479692549,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 16012,
+          "value": 7064.044106984511,
+          "p10": 7060.379757732153,
+          "p90": 7075.317796040326,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 10,
+      "model": "qwen3.8-27b-mtp",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-16",
+      "finished_at": "2026-09-17",
+      "adopted_from_date": null,
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 80.69203693375472,
+      "categories": {
+        "knowledge": 71.98484848484848,
+        "math": 88.57777777777778,
+        "code": 83.20552147239263,
+        "german": 79.0
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 80.33333333333333,
+          "stderr": 2.298675559401019,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 63.63636363636363,
+          "stderr": 3.4273086529999337,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 91.2,
+          "stderr": 1.7953084777052886,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 83.33333333333334,
+          "stderr": 3.0530971060304655,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 91.41104294478528,
+          "stderr": 2.2014662933817504,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 75.0,
+          "stderr": 3.0695456590127175,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 79.0,
+          "stderr": 2.355524354210245,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 15.103805505442727,
+          "p10": 14.29388046074622,
+          "p90": 15.614755696822689,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 1836.5705961540752,
+          "p10": 1822.4245126148585,
+          "p90": 1852.5270695868603,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4024,
+          "value": 2550.5619640606633,
+          "p10": 2342.9211020025855,
+          "p90": 2556.148618285338,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 16001,
+          "value": 1734.2630471793032,
+          "p10": 1676.5056173024004,
+          "p90": 1735.1294735680774,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 288.03684492595494,
+          "p10": 285.3419618215412,
+          "p90": 290.27276262640953,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4024,
+          "value": 1577.6915270835161,
+          "p10": 1574.2433857172728,
+          "p90": 1726.2721830047667,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 16001,
+          "value": 9226.397359976545,
+          "p10": 9221.79028405808,
+          "p90": 9551.418171450496,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 11,
+      "model": "qwen3.5-9b-heretic-v2-mtp",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-16",
+      "finished_at": "2026-09-17",
+      "adopted_from_date": null,
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 73.63538219619508,
+      "categories": {
+        "knowledge": 67.2070707070707,
+        "math": 80.53333333333333,
+        "code": 77.46779141104295,
+        "german": 69.33333333333334
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 76.33333333333333,
+          "stderr": 2.4580463430538737,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 58.080808080808076,
+          "stderr": 3.5155207286704204,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 76.8,
+          "stderr": 2.67500703748652,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 88.0,
+          "stderr": 2.6621886338401435,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 83.43558282208589,
+          "stderr": 2.9208296231259103,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 71.5,
+          "stderr": 3.1999921482315727,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 69.33333333333334,
+          "stderr": 2.6666666666666665,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 31.530246994028012,
+          "p10": 30.996760502842868,
+          "p90": 31.970842472199518,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 4737.205615904939,
+          "p10": 4672.451505609171,
+          "p90": 4758.431814142912,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4024,
+          "value": 6290.925546819461,
+          "p10": 6277.162900690696,
+          "p90": 6307.3356574246955,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 16001,
+          "value": 5096.428585224607,
+          "p10": 5083.639418179652,
+          "p90": 5104.59062554422,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 111.66920815594494,
+          "p10": 111.17180613800883,
+          "p90": 113.14096082933247,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4024,
+          "value": 639.6515059750527,
+          "p10": 637.9873093217611,
+          "p90": 641.0549090243876,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 16001,
+          "value": 3139.6496060770005,
+          "p10": 3134.631334710866,
+          "p90": 3147.558627743274,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    }
+  ]
+}

runs/10.html

@@ -0,0 +1,181 @@
+<!doctype html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.8-27b-mtp – 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">Übersicht</a>
+    <a href="../methodik.html">Methodik</a>
+    <a href="../results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v2-quick">← Übersicht</a></p>
+<h1>qwen3.8-27b-mtp</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-16 – 2026-09-17</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inferenz-Software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Messweg</dt><dd>direkt am Inferenz-Server</dd>
+  <dt>Gesamt</dt><dd class="strong">80.7</dd>
+</dl>
+
+<h2>Kategorien</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Wissen &amp; Reasoning</td><td class="num">72.0</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">88.6</td></tr>
+<tr><td>Code</td><td class="num">83.2</td></tr>
+<tr><td>Deutsch</td><td class="num">79.0</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">80.3</td>
+  <td class="num">± 2.3</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">63.6</td>
+  <td class="num">± 3.4</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td class="num">91.2</td>
+  <td class="num">± 1.8</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>ifeval</td>
+  <td class="num">83.3</td>
+  <td class="num">± 3.1</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>humaneval_plus_gen</td>
+  <td class="num">91.4</td>
+  <td class="num">± 2.2</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">75.0</td>
+  <td class="num">± 3.1</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">79.0</td>
+  <td class="num">± 2.4</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± ist der Standardfehler; das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.</p>
+
+<h2>Geschwindigkeit</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Messgröße</th><th class="num">Kontext (Ziel)</th><th class="num">Kontext (echt)</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>Generierung [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">15.1</td>
+  <td class="num">14.3</td>
+  <td class="num">15.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">1836.6</td>
+  <td class="num">1822.4</td>
+  <td class="num">1852.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">2550.6</td>
+  <td class="num">2342.9</td>
+  <td class="num">2556.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">1734.3</td>
+  <td class="num">1676.5</td>
+  <td class="num">1735.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">288.0</td>
+  <td class="num">285.3</td>
+  <td class="num">290.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">1577.7</td>
+  <td class="num">1574.2</td>
+  <td class="num">1726.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">9226.4</td>
+  <td class="num">9221.8</td>
+  <td class="num">9551.4</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/11.html

@@ -0,0 +1,181 @@
+<!doctype html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.5-9b-heretic-v2-mtp – 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">Übersicht</a>
+    <a href="../methodik.html">Methodik</a>
+    <a href="../results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v2-quick">← Übersicht</a></p>
+<h1>qwen3.5-9b-heretic-v2-mtp</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-16 – 2026-09-17</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inferenz-Software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Messweg</dt><dd>direkt am Inferenz-Server</dd>
+  <dt>Gesamt</dt><dd class="strong">73.6</dd>
+</dl>
+
+<h2>Kategorien</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Wissen &amp; Reasoning</td><td class="num">67.2</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">80.5</td></tr>
+<tr><td>Code</td><td class="num">77.5</td></tr>
+<tr><td>Deutsch</td><td class="num">69.3</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">76.3</td>
+  <td class="num">± 2.5</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">58.1</td>
+  <td class="num">± 3.5</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td class="num">76.8</td>
+  <td class="num">± 2.7</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>ifeval</td>
+  <td class="num">88.0</td>
+  <td class="num">± 2.7</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>humaneval_plus_gen</td>
+  <td class="num">83.4</td>
+  <td class="num">± 2.9</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">71.5</td>
+  <td class="num">± 3.2</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">69.3</td>
+  <td class="num">± 2.7</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± ist der Standardfehler; das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.</p>
+
+<h2>Geschwindigkeit</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Messgröße</th><th class="num">Kontext (Ziel)</th><th class="num">Kontext (echt)</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>Generierung [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">31.5</td>
+  <td class="num">31.0</td>
+  <td class="num">32.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">4737.2</td>
+  <td class="num">4672.5</td>
+  <td class="num">4758.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">6290.9</td>
+  <td class="num">6277.2</td>
+  <td class="num">6307.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">5096.4</td>
+  <td class="num">5083.6</td>
+  <td class="num">5104.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">111.7</td>
+  <td class="num">111.2</td>
+  <td class="num">113.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">639.7</td>
+  <td class="num">638.0</td>
+  <td class="num">641.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">3139.6</td>
+  <td class="num">3134.6</td>
+  <td class="num">3147.6</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/6.html

@@ -0,0 +1,181 @@
+<!doctype html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>nemotron-3.5-lightning-dspark – 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">Übersicht</a>
+    <a href="../methodik.html">Methodik</a>
+    <a href="../results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v2-quick">← Übersicht</a></p>
+<h1>nemotron-3.5-lightning-dspark</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-15</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inferenz-Software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Messweg</dt><dd>direkt am Inferenz-Server</dd>
+  <dt>Gesamt</dt><dd class="strong">76.3</dd>
+</dl>
+
+<h2>Kategorien</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Wissen &amp; Reasoning</td><td class="num">64.0</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">86.7</td></tr>
+<tr><td>Code</td><td class="num">82.6</td></tr>
+<tr><td>Deutsch</td><td class="num">72.0</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">71.0</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">57.1</td>
+  <td class="num">± 3.5</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td class="num">90.4</td>
+  <td class="num">± 1.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>ifeval</td>
+  <td class="num">79.3</td>
+  <td class="num">± 3.3</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>humaneval_plus_gen</td>
+  <td class="num">87.7</td>
+  <td class="num">± 2.6</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">77.5</td>
+  <td class="num">± 3.0</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">72.0</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± ist der Standardfehler; das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.</p>
+
+<h2>Geschwindigkeit</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Messgröße</th><th class="num">Kontext (Ziel)</th><th class="num">Kontext (echt)</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>Generierung [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">83.8</td>
+  <td class="num">80.7</td>
+  <td class="num">87.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">528</td>
+  <td class="num">3738.2</td>
+  <td class="num">2851.3</td>
+  <td class="num">3772.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4010</td>
+  <td class="num">7236.8</td>
+  <td class="num">6862.4</td>
+  <td class="num">7270.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15960</td>
+  <td class="num">7177.3</td>
+  <td class="num">7129.3</td>
+  <td class="num">7180.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">528</td>
+  <td class="num">141.2</td>
+  <td class="num">140.0</td>
+  <td class="num">185.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4010</td>
+  <td class="num">554.1</td>
+  <td class="num">551.6</td>
+  <td class="num">585.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15960</td>
+  <td class="num">2223.7</td>
+  <td class="num">2222.8</td>
+  <td class="num">2238.7</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/7.html

@@ -0,0 +1,181 @@
+<!doctype html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>gemma-4-12b-it-heretic – 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">Übersicht</a>
+    <a href="../methodik.html">Methodik</a>
+    <a href="../results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v2-quick">← Übersicht</a></p>
+<h1>gemma-4-12b-it-heretic</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-15</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inferenz-Software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Messweg</dt><dd>direkt am Inferenz-Server</dd>
+  <dt>Gesamt</dt><dd class="strong">76.9</dd>
+</dl>
+
+<h2>Kategorien</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Wissen &amp; Reasoning</td><td class="num">66.4</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">86.6</td></tr>
+<tr><td>Code</td><td class="num">84.7</td></tr>
+<tr><td>Deutsch</td><td class="num">70.0</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &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>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">56.1</td>
+  <td class="num">± 3.5</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td class="num">85.6</td>
+  <td class="num">± 2.2</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</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>humaneval_plus_gen</td>
+  <td class="num">94.5</td>
+  <td class="num">± 1.8</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">75.0</td>
+  <td class="num">± 3.1</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">70.0</td>
+  <td class="num">± 2.7</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± ist der Standardfehler; das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.</p>
+
+<h2>Geschwindigkeit</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Messgröße</th><th class="num">Kontext (Ziel)</th><th class="num">Kontext (echt)</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>Generierung [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">21.4</td>
+  <td class="num">21.4</td>
+  <td class="num">21.4</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">528</td>
+  <td class="num">5070.5</td>
+  <td class="num">4947.7</td>
+  <td class="num">5118.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4014</td>
+  <td class="num">5258.0</td>
+  <td class="num">5221.7</td>
+  <td class="num">5293.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15985</td>
+  <td class="num">3514.1</td>
+  <td class="num">3505.2</td>
+  <td class="num">3518.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">528</td>
+  <td class="num">103.9</td>
+  <td class="num">103.2</td>
+  <td class="num">106.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4014</td>
+  <td class="num">763.4</td>
+  <td class="num">758.2</td>
+  <td class="num">768.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15985</td>
+  <td class="num">4548.8</td>
+  <td class="num">4543.1</td>
+  <td class="num">4560.4</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/8.html

@@ -0,0 +1,181 @@
+<!doctype html>
+<html lang="de">
+<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">Übersicht</a>
+    <a href="../methodik.html">Methodik</a>
+    <a href="../results.json">Daten (JSON)</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v2-quick">← Übersicht</a></p>
+<h1>qwen3.8-flash-next-ple4</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-15 – 2026-09-16</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inferenz-Software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Messweg</dt><dd>direkt am Inferenz-Server</dd>
+  <dt>Gesamt</dt><dd class="strong">87.7</dd>
+</dl>
+
+<h2>Kategorien</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Wissen &amp; Reasoning</td><td class="num">81.8</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">95.3</td></tr>
+<tr><td>Code</td><td class="num">87.9</td></tr>
+<tr><td>Deutsch</td><td class="num">85.7</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Kategorie</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">86.3</td>
+  <td class="num">± 2.0</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">77.3</td>
+  <td class="num">± 3.0</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>math_hard_gen</td>
+  <td class="num">97.6</td>
+  <td class="num">± 1.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</td>
+  <td>ifeval</td>
+  <td class="num">90.7</td>
+  <td class="num">± 2.4</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>humaneval_plus_gen</td>
+  <td class="num">96.3</td>
+  <td class="num">± 1.5</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">79.5</td>
+  <td class="num">± 2.9</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">85.7</td>
+  <td class="num">± 2.0</td>
+  <td class="num">300</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± ist der Standardfehler; das 95-%-Intervall reicht etwa 1,96 Standardfehler um den Score.</p>
+
+<h2>Geschwindigkeit</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Messgröße</th><th class="num">Kontext (Ziel)</th><th class="num">Kontext (echt)</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>Generierung [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">26.4</td>
+  <td class="num">26.3</td>
+  <td class="num">26.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">1051.1</td>
+  <td class="num">1048.4</td>
+  <td class="num">1056.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4029</td>
+  <td class="num">2041.5</td>
+  <td class="num">2032.9</td>
+  <td class="num">2044.5</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16012</td>
+  <td class="num">2266.7</td>
+  <td class="num">2263.1</td>
+  <td class="num">2267.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">503.3</td>
+  <td class="num">500.3</td>
+  <td class="num">504.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4029</td>
+  <td class="num">1973.6</td>
+  <td class="num">1970.7</td>
+  <td class="num">1982.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16012</td>
+  <td class="num">7064.0</td>
+  <td class="num">7060.4</td>
+  <td class="num">7075.3</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

style.css

@@ -1,28 +1,151 @@
-body {
-	padding: 2rem;
-	font-family: -apple-system, BlinkMacSystemFont, "Arial", sans-serif;
+:root {
+  color-scheme: light;
+  --page: #f9f9f7;
+  --surface: #fcfcfb;
+  --ink: #0b0b0b;
+  --ink-2: #52514e;
+  --muted: #898781;
+  --grid: #e1e0d9;
+  --axis: #c3c2b7;
+  --border: rgba(11, 11, 11, 0.10);
+  --accent: #2a78d6;
+  --series-1: #2a78d6;
+  --series-2: #eb6834;
+  --series-3: #1baf7a;
+  --warn: #fab219;
+  --warn-ink: #7a5200;
+  --danger: #c62828;
+  --ok: #1b7f4b;
+  --medal-gold: #b8860b;
+  --medal-gold-ink: #7a5a00;
+  --medal-silver: #8a8f98;
+  --medal-silver-ink: #52565c;
+  --medal-bronze: #a9673a;
+  --medal-bronze-ink: #7a4522;
 }
-
-h1 {
-	font-size: 16px;
-	margin-top: 0;
+@media (prefers-color-scheme: dark) {
+  :root:not([data-theme="light"]) {
+    color-scheme: dark;
+    --page: #0d0d0d;
+    --surface: #1a1a19;
+    --ink: #ffffff;
+    --ink-2: #c3c2b7;
+    --muted: #898781;
+    --grid: #2c2c2a;
+    --axis: #383835;
+    --border: rgba(255, 255, 255, 0.10);
+    --accent: #3987e5;
+    --series-1: #3987e5;
+    --series-2: #d95926;
+    --series-3: #199e70;
+    --warn-ink: #fab219;
+    --danger: #ef5350;
+    --ok: #4cc38a;
+    --medal-gold-ink: #e8c547;
+    --medal-silver-ink: #c7ccd1;
+    --medal-bronze-ink: #d99a6c;
+  }
 }
+* { box-sizing: border-box; }
+body { margin: 0; background: var(--page); color: var(--ink); font: 14px/1.45 system-ui, -apple-system, "Segoe UI", sans-serif; }
 
-p {
-	color: rgb(107, 114, 128);
-	font-size: 15px;
-	margin-bottom: 10px;
-	margin-top: 5px;
+/* Links: Inhalt in Textfarbe, Akzent erst beim Überfahren; Nebenlinks klein und grau */
+a { color: var(--ink); text-decoration: none; text-underline-offset: 2px; }
+a:hover { color: var(--accent); text-decoration: underline; }
+.sub a, a.secondary { font-size: 12px; color: var(--muted); }
+.sub a:hover, a.secondary:hover { color: var(--accent); }
+a:focus-visible, button:focus-visible, input:focus-visible, select:focus-visible, summary:focus-visible {
+  outline: 2px solid var(--accent); outline-offset: 2px; border-radius: 4px;
 }
 
-.card {
-	max-width: 620px;
-	margin: 0 auto;
-	padding: 16px;
-	border: 1px solid lightgray;
-	border-radius: 16px;
-}
+.top { display: flex; flex-wrap: wrap; gap: 8px 16px; align-items: center; justify-content: space-between; padding: 8px 16px; background: var(--surface); border-bottom: 1px solid var(--border); }
+.brand { font-weight: 600; color: var(--ink); }
+.brand:hover { text-decoration: none; }
+nav { display: flex; flex-wrap: wrap; gap: 4px; }
+nav a { padding: 6px 10px; color: var(--ink-2); border-bottom: 2px solid transparent; }
+nav a:hover { color: var(--ink); text-decoration: none; border-bottom-color: var(--border); }
+nav a.active { color: var(--ink); font-weight: 600; border-bottom-color: var(--accent); }
+main { max-width: 1400px; margin: 0 auto; padding: 16px; }
+h1 { font-size: 20px; margin: 8px 0 16px; overflow-wrap: anywhere; }
+h2 { font-size: 16px; margin: 24px 0 8px; }
+h3 { font-size: 13px; margin: 8px 0 4px; color: var(--ink-2); }
+.toolbar { display: flex; flex-wrap: wrap; gap: 12px; align-items: center; justify-content: space-between; margin: 12px 0; }
+.segmented a { display: inline-block; padding: 6px 12px; border: 1px solid var(--border); color: var(--ink-2); }
+.segmented a:hover { text-decoration: none; color: var(--ink); }
+.segmented a:first-child { border-radius: 6px 0 0 6px; }
+.segmented a:last-child { border-radius: 0 6px 6px 0; }
+.segmented a.active { background: var(--surface); color: var(--ink); font-weight: 600; }
 
-.card p:last-child {
-	margin-bottom: 0;
-}
+button { font: inherit; padding: 6px 12px; border-radius: 6px; border: 1px solid var(--border); background: var(--surface); color: var(--ink); cursor: pointer; }
+button:disabled { opacity: 0.5; cursor: default; }
+button.primary { background: var(--accent); border-color: var(--accent); color: #fff; }
+button.danger { color: var(--danger); border-color: var(--danger); }
+button.danger:hover:not(:disabled) { background: var(--danger); color: #fff; }
+.actions { display: flex; flex-wrap: wrap; gap: 8px; align-items: center; margin: 12px 0; }
+.actions form, .inline-form { display: inline; margin: 0; }
+input, select { font: inherit; padding: 6px 8px; border-radius: 6px; border: 1px solid var(--border); background: var(--page); color: var(--ink); min-width: 0; }
+
+.table-wrap { overflow-x: auto; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
+table { border-collapse: collapse; width: 100%; }
+th, td { padding: 8px 10px; border-bottom: 1px solid var(--grid); text-align: left; vertical-align: top; white-space: nowrap; }
+th { font-size: 12px; font-weight: 600; color: var(--ink-2); white-space: normal; vertical-align: bottom; }
+th[data-sort] { cursor: pointer; user-select: none; }
+th[aria-sort="ascending"]::after { content: " ▲"; }
+th[aria-sort="descending"]::after { content: " ▼"; }
+.num { text-align: right; font-variant-numeric: tabular-nums; }
+.strong { font-weight: 600; }
+.sub { font-size: 12px; color: var(--muted); }
+.wrap { white-space: normal; max-width: 420px; }
+.muted { color: var(--muted); }
+.badge { display: inline-block; margin-top: 4px; padding: 1px 6px; border: 1px solid var(--border); border-radius: 10px; font-size: 11px; color: var(--ink-2); }
+.badge.warn, .warn-text { color: var(--warn-ink); border-color: var(--warn); }
+.hint { font-size: 12px; color: var(--ink-2); }
+.flash { padding: 8px 12px; background: var(--surface); border: 1px solid var(--border); border-left: 4px solid var(--accent); border-radius: 6px; }
+.empty { padding: 24px; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
+.visually-hidden { position: absolute; width: 1px; height: 1px; overflow: hidden; clip: rect(0 0 0 0); }
+
+.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-box { position: relative; width: 100%; }
+.meta { display: grid; grid-template-columns: max-content 1fr; gap: 4px 16px; padding: 12px; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
+.meta dt { color: var(--ink-2); }
+.meta dd { margin: 0; overflow-wrap: anywhere; }
+details { margin: 12px 0; }
+pre { margin: 0; padding: 8px; max-height: 360px; overflow: auto; white-space: pre-wrap; word-break: break-word; background: var(--page); border-radius: 6px; }
+pre.log { max-height: 320px; font-size: 12px; }
+.answer { margin: 12px 0; padding: 12px; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
+.pair { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(100%, 320px), 1fr)); gap: 12px; margin-top: 8px; }
+.pager a { margin-right: 16px; color: var(--accent); }
+.compare td.best { font-weight: 600; }
+.compare td.greyed { color: var(--muted); }
+.compare tr.total td { font-weight: 600; background: var(--page); }
+.medal { display: inline-block; margin-left: 4px; padding: 0 5px; border: 1px solid var(--border); border-radius: 10px; font-size: 10px; font-weight: 600; line-height: 14px; }
+.medal.gold { color: var(--medal-gold-ink); border-color: var(--medal-gold); }
+.medal.silver { color: var(--medal-silver-ink); border-color: var(--medal-silver); }
+.medal.bronze { color: var(--medal-bronze-ink); border-color: var(--medal-bronze); }
+
+/* Seite Betrieb */
+.status { padding: 8px 12px; border: 1px solid var(--border); border-radius: 6px; background: var(--surface); }
+.status.ok::before { content: "● "; color: var(--ok); }
+.status.bad { border-color: var(--danger); }
+.status.bad::before { content: "▲ "; color: var(--danger); }
+.cards { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(100%, 360px), 1fr)); gap: 12px; }
+.card { padding: 12px; background: var(--surface); border: 1px solid var(--border); border-radius: 8px; }
+.card-head { display: flex; flex-wrap: wrap; gap: 8px; align-items: baseline; overflow-wrap: anywhere; }
+.card p { margin: 6px 0; }
+progress { width: 100%; height: 8px; accent-color: var(--accent); }
+.form-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(100%, 220px), 1fr)); gap: 12px; }
+.form-grid .wide { grid-column: 1 / -1; }
+.form-grid label { display: flex; flex-direction: column; gap: 4px; font-size: 12px; color: var(--ink-2); }
+.form-grid label.inline { flex-direction: row; align-items: center; gap: 6px; font-size: 14px; color: var(--ink); }
+fieldset { display: flex; flex-wrap: wrap; gap: 8px 20px; margin: 0; border: 1px solid var(--border); border-radius: 6px; }
+legend { font-size: 12px; color: var(--ink-2); }
+
+/* Ergänzungen für die veröffentlichte statische Seite (bench publish) */
+.lead { max-width: 820px; color: var(--ink-2); }
+.prose { max-width: 820px; }
+.prose p, .prose li { color: var(--ink-2); }
+section + section { margin-top: 24px; }
+.page-foot { margin-top: 32px; padding-top: 12px; border-top: 1px solid var(--border); font-size: 12px; color: var(--muted); }
+p.segmented { display: flex; flex-wrap: wrap; margin: 12px 0; }