Ergebnisse aktualisiert (14 Läufe)

index.html

@@ -18,6 +18,133 @@
 <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>
+<p class="segmented" aria-label="Abschnitte"><a href="#v3-quick">v2b · quick</a><a href="#v2-quick">v2 · quick</a></p>
+<section id="v3-quick">
+<h2>Suite v2b · 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</th><th data-sort="num" class="num">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/15.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="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 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 strong" data-value="75.5" title="livecodebench_gen: 75.5 ± 3.0 (n=200)&#10;">75.5</td>
+  <td class="num strong" data-value="81.13708513708514" title="mmlu_prox_de: 85.7 ± 2.0 (n=300)&#10;include_de_gen: 72.1 ± 3.6 (n=154)&#10;">81.1</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="2026-09-17">2026-09-17<div class="sub">übernommen aus Messung vom 2026-09-15</div></td>
+</tr>
+<tr>
+  <td data-value="gemma-4-12b-it">
+    <a href="runs/21.html" class="strong">gemma-4-12b-it</a>
+    <div class="sub">Q8 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="65.82462114883167">65.8</td>
+  <td class="num" data-value="67.48821548821549" title="mmlu_pro: 79.3 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 63.1 ± 3.4 (n=198)&#10;musr_gen: 60.0 ± 3.1 (n=250)&#10;">67.5</td>
+  <td class="num" data-value="60.8" title="omni_math_gen: 60.8 ± 3.1 (n=250)&#10;">60.8</td>
+  <td class="num strong" data-value="51.318295739348365" title="ifbench: 38.3 ± 2.8 (n=300)&#10;ifeval: 90.7 ± 2.4 (n=150)&#10;">51.3</td>
+  <td class="num" data-value="75.0" title="livecodebench_gen: 75.0 ± 3.1 (n=200)&#10;">75.0</td>
+  <td class="num" data-value="74.51659451659451" title="mmlu_prox_de: 78.3 ± 2.4 (n=300)&#10;include_de_gen: 66.9 ± 3.8 (n=154)&#10;">74.5</td>
+  <td class="num" data-value="16.51787775562755">17</td>
+  <td class="num" data-value="1359.0489942173106">1359</td>
+  <td class="num" data-value="2952.064286917448">2952</td>
+  <td data-value="2026-09-18">2026-09-18<div class="sub">übernommen aus Messung vom 2026-09-17</div></td>
+</tr>
+<tr>
+  <td data-value="gemma-4-12b-it-heretic">
+    <a href="runs/22.html" class="strong">gemma-4-12b-it-heretic</a>
+    <div class="sub">Q8 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="65.82311637680058">65.8</td>
+  <td class="num" data-value="68.44713804713804" title="mmlu_pro: 80.0 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 64.1 ± 3.4 (n=198)&#10;musr_gen: 61.2 ± 3.1 (n=250)&#10;">68.4</td>
+  <td class="num" data-value="59.199999999999996" title="omni_math_gen: 59.2 ± 3.1 (n=250)&#10;">59.2</td>
+  <td class="num" data-value="49.97493734335839" title="ifbench: 35.7 ± 2.8 (n=300)&#10;ifeval: 93.3 ± 2.0 (n=150)&#10;">50.0</td>
+  <td class="num" data-value="75.0" title="livecodebench_gen: 75.0 ± 3.1 (n=200)&#10;">75.0</td>
+  <td class="num" data-value="76.4935064935065" title="mmlu_prox_de: 80.0 ± 2.3 (n=300)&#10;include_de_gen: 69.5 ± 3.7 (n=154)&#10;">76.5</td>
+  <td class="num" data-value="16.51271607105954">17</td>
+  <td class="num" data-value="1358.107397713914">1358</td>
+  <td class="num" data-value="2954.110997961834">2954</td>
+  <td data-value="2026-09-19">2026-09-19<div class="sub">übernommen aus Messung vom 2026-09-17</div></td>
+</tr>
+<tr>
+  <td data-value="qwen3.8-27b-mtp">
+    <a href="runs/20.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="64.20301055669476">64.2</td>
+  <td class="num" data-value="69.85656565656565" title="mmlu_pro: 80.3 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 63.6 ± 3.4 (n=198)&#10;musr_gen: 65.6 ± 3.0 (n=250)&#10;">69.9</td>
+  <td class="num" data-value="60.4" title="omni_math_gen: 60.4 ± 3.1 (n=250)&#10;">60.4</td>
+  <td class="num" data-value="49.498746867167924" title="ifbench: 38.3 ± 2.8 (n=300)&#10;ifeval: 83.3 ± 3.1 (n=150)&#10;">49.5</td>
+  <td class="num" data-value="65.0" title="livecodebench_gen: 65.0 ± 3.4 (n=200)&#10;">65.0</td>
+  <td class="num" data-value="76.25974025974027" title="mmlu_prox_de: 79.0 ± 2.4 (n=300)&#10;include_de_gen: 70.8 ± 3.7 (n=154)&#10;">76.3</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="2026-09-18">2026-09-18<div class="sub">übernommen aus Messung vom 2026-09-16</div></td>
+</tr>
+<tr>
+  <td data-value="gemma-4-12b-it-heretic">
+    <a href="runs/17.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="57.7166917293233">57.7</td>
+  <td class="num" data-value="63.70909090909091" title="mmlu_pro: 76.7 ± 2.4 (n=300)&#10;gpqa_diamond_gen: 56.1 ± 3.5 (n=198)&#10;musr_gen: 58.4 ± 3.1 (n=250)&#10;">63.7</td>
+  <td class="num" data-value="48.4" title="omni_math_gen: 48.4 ± 3.2 (n=250)&#10;">48.4</td>
+  <td class="num" data-value="46.81203007518797" title="ifbench: 33.0 ± 2.7 (n=300)&#10;ifeval: 88.7 ± 2.6 (n=150)&#10;">46.8</td>
+  <td class="num" data-value="62.0" title="livecodebench_gen: 62.0 ± 3.4 (n=200)&#10;">62.0</td>
+  <td class="num" data-value="67.66233766233766" title="mmlu_prox_de: 70.0 ± 2.7 (n=300)&#10;include_de_gen: 63.0 ± 3.9 (n=154)&#10;">67.7</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="2026-09-18">2026-09-18<div class="sub">übernommen aus Messung vom 2026-09-15</div></td>
+</tr>
+<tr>
+  <td data-value="nemotron-3.5-lightning-dspark">
+    <a href="runs/16.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="53.13354497354497">53.1</td>
+  <td class="num" data-value="55.756902356902366" title="mmlu_pro: 71.0 ± 2.6 (n=300)&#10;gpqa_diamond_gen: 57.1 ± 3.5 (n=198)&#10;musr_gen: 39.2 ± 3.1 (n=250)&#10;">55.8</td>
+  <td class="num" data-value="40.400000000000006" title="omni_math_gen: 40.4 ± 3.1 (n=250)&#10;">40.4</td>
+  <td class="num" data-value="46.0" title="ifbench: 35.0 ± 2.8 (n=300)&#10;ifeval: 79.3 ± 3.3 (n=150)&#10;">46.0</td>
+  <td class="num" data-value="53.0" title="livecodebench_gen: 53.0 ± 3.5 (n=200)&#10;">53.0</td>
+  <td class="num" data-value="70.5108225108225" title="mmlu_prox_de: 72.0 ± 2.6 (n=300)&#10;include_de_gen: 67.5 ± 3.8 (n=154)&#10;">70.5</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="2026-09-17">2026-09-17<div class="sub">übernommen aus Messung vom 2026-09-15</div></td>
+</tr>
+<tr>
+  <td data-value="qwen3.5-9b-heretic-v2-mtp">
+    <a href="runs/19.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="51.662634870003295">51.7</td>
+  <td class="num" data-value="65.07138047138046" title="mmlu_pro: 76.3 ± 2.5 (n=300)&#10;gpqa_diamond_gen: 58.1 ± 3.5 (n=198)&#10;musr_gen: 60.8 ± 3.1 (n=250)&#10;">65.1</td>
+  <td class="num" data-value="35.6" title="omni_math_gen: 35.6 ± 3.0 (n=250)&#10;">35.6</td>
+  <td class="num" data-value="44.140350877192986" title="ifbench: 29.7 ± 2.6 (n=300)&#10;ifeval: 88.0 ± 2.7 (n=150)&#10;">44.1</td>
+  <td class="num" data-value="46.5" title="livecodebench_gen: 46.5 ± 3.5 (n=200)&#10;">46.5</td>
+  <td class="num" data-value="67.00144300144301" title="mmlu_prox_de: 69.3 ± 2.7 (n=300)&#10;include_de_gen: 62.3 ± 3.9 (n=154)&#10;">67.0</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="2026-09-18">2026-09-18<div class="sub">übernommen aus Messung vom 2026-09-16</div></td>
+</tr>
+</tbody>
+</table>
+</div>
+</section>
 <section id="v2-quick">
 <h2>Suite v2 · Profil quick</h2>
 <div class="table-wrap">
@@ -43,6 +170,36 @@
   <td class="num" data-value="1973.583644023165">1974</td>
   <td data-value="2026-09-15">2026-09-15</td>
 </tr>
+<tr>
+  <td data-value="gemma-4-12b-it-heretic">
+    <a href="runs/14.html" class="strong">gemma-4-12b-it-heretic</a>
+    <div class="sub">Q8 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="83.09445761293921">83.1</td>
+  <td class="num" data-value="72.07070707070707" title="mmlu_pro: 80.0 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 64.1 ± 3.4 (n=198)&#10;">72.1</td>
+  <td class="num" data-value="93.51111111111112" title="math_hard_gen: 93.6 ± 1.6 (n=250)&#10;ifeval: 93.3 ± 2.0 (n=150)&#10;">93.5</td>
+  <td class="num" data-value="86.79601226993864" title="humaneval_plus_gen: 95.1 ± 1.7 (n=163)&#10;mbpp_plus_gen: 78.5 ± 2.9 (n=200)&#10;">86.8</td>
+  <td class="num" data-value="80.0" title="mmlu_prox_de: 80.0 ± 2.3 (n=300)&#10;">80.0</td>
+  <td class="num" data-value="16.51271607105954">17</td>
+  <td class="num" data-value="1358.107397713914">1358</td>
+  <td class="num" data-value="2954.110997961834">2954</td>
+  <td data-value="2026-09-17">2026-09-17</td>
+</tr>
+<tr>
+  <td data-value="gemma-4-12b-it">
+    <a href="runs/12.html" class="strong">gemma-4-12b-it</a>
+    <div class="sub">Q8 · Reasoning off · direkt · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="82.04597276445436">82.0</td>
+  <td class="num" data-value="71.23232323232322" title="mmlu_pro: 79.3 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 63.1 ± 3.4 (n=198)&#10;">71.2</td>
+  <td class="num" data-value="91.82222222222224" title="math_hard_gen: 92.4 ± 1.7 (n=250)&#10;ifeval: 90.7 ± 2.4 (n=150)&#10;">91.8</td>
+  <td class="num" data-value="86.79601226993864" title="humaneval_plus_gen: 95.1 ± 1.7 (n=163)&#10;mbpp_plus_gen: 78.5 ± 2.9 (n=200)&#10;">86.8</td>
+  <td class="num" data-value="78.33333333333333" title="mmlu_prox_de: 78.3 ± 2.4 (n=300)&#10;">78.3</td>
+  <td class="num" data-value="16.51787775562755">17</td>
+  <td class="num" data-value="1359.0489942173106">1359</td>
+  <td class="num" data-value="2952.064286917448">2952</td>
+  <td data-value="2026-09-17">2026-09-17</td>
+</tr>
 <tr>
   <td data-value="qwen3.8-27b-mtp">
     <a href="runs/10.html" class="strong">qwen3.8-27b-mtp</a>
@@ -108,7 +265,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="app.js"></script>
 </body>

methodik.html

@@ -38,6 +38,86 @@ sind daher etwas langsamer. Der Messweg steht auf der Seite des Laufs. Geschwind
 Inferenz-Software ab.</p>
 </section>
 
+<h2>Suite v2b</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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td>exact_match,letter</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen<div class="sub wrap">Omni-MATH-Rule ist die regelbasiert prüfbare Teilmenge von Omni-MATH und systematisch leichter als das volle Omni-MATH. Die Werte sind deshalb nicht mit veröffentlichten Omni-MATH-Zahlen vergleichbar. Der Datensatz ist seit Oktober 2024 öffentlich und liegt vor den Trainings-Cutoffs aller gemessenen Modelle; das betrifft alle Modelle gleich.</div></td>
+  <td>exact_match,boxed</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td>prompt_level_strict_acc,none</td>
+  <td>Text</td>
+  <td class="num">1</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifeval</td>
+  <td>prompt_level_strict_acc,none</td>
+  <td>Text</td>
+  <td class="num">0.33</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen<div class="sub wrap">Kontaminiert, und zwar ungleich: LiveCodeBench v6 enthält Aufgaben von Mai 2023 bis April 2025, ein neueres Release gibt es nicht. Die Trainings-Cutoffs der gemessenen Modelle liegen teils dahinter (Stand September 2026: Gemma 4 Januar 2025, Nemotron 3.5 Lightning September 2025, Qwen3.8 nicht veröffentlicht). Ein Modell mit späterem Cutoff kann einen größeren Teil der Aufgaben im Training gesehen haben; Abstände zwischen Modellen sind hier nur eingeschränkt aussagekräftig. Ausgaben werden exakt wie im offiziellen LiveCodeBench verglichen: 12 Aufgaben mit Fließkomma-Ausgabe sind dadurch für alle Modelle praktisch unlösbar (auch in den veröffentlichten LiveCodeBench-Zahlen); das senkt den erreichbaren Höchstwert für alle gleich. Liste und Abgleich mit dem offiziellen Auswerter im Kalibrierungsbericht (docs/superpowers/reports/2026-09-16-lcb-kalibrierung.md). Profil full lädt den vollen Datensatz und braucht dafür rund 9 GB RAM.</div></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>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen<div class="sub wrap">Nur 154 Aufgaben (Test und Validation von INCLUDE-German, alle werden gemessen). Das 95-%-Intervall beträgt rund ±8 Punkte; kleinere Abstände zwischen Modellen sind Rauschen. Deshalb halbes Gewicht neben mmlu_prox_de.</div></td>
+  <td>exact_match,letter</td>
+  <td>Text</td>
+  <td class="num">0.5</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
 <h2>Suite v2</h2>
 <div class="table-wrap">
 <table>
@@ -102,7 +182,7 @@ Inferenz-Software ab.</p>
 </tbody>
 </table>
 </div>
-<footer class="page-foot">Stand 2026-09-17 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="app.js"></script>
 </body>

results.json

@@ -1,8 +1,110 @@
 {
   "schema_version": 1,
   "generator": "bench publish",
-  "generated_at": "2026-09-17",
+  "generated_at": "2026-09-20",
   "suites": [
+    {
+      "version": 3,
+      "label": "v2b",
+      "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
+            },
+            {
+              "name": "musr_gen",
+              "metric": "exact_match,letter",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 250
+            }
+          ]
+        },
+        {
+          "key": "math",
+          "label": "Mathe",
+          "tasks": [
+            {
+              "name": "omni_math_gen",
+              "metric": "exact_match,boxed",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 250,
+              "note": "Omni-MATH-Rule ist die regelbasiert prüfbare Teilmenge von Omni-MATH und systematisch leichter als das volle Omni-MATH. Die Werte sind deshalb nicht mit veröffentlichten Omni-MATH-Zahlen vergleichbar. Der Datensatz ist seit Oktober 2024 öffentlich und liegt vor den Trainings-Cutoffs aller gemessenen Modelle; das betrifft alle Modelle gleich."
+            }
+          ]
+        },
+        {
+          "key": "instructions",
+          "label": "Instruktionen",
+          "tasks": [
+            {
+              "name": "ifbench",
+              "metric": "prompt_level_strict_acc,none",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 300
+            },
+            {
+              "name": "ifeval",
+              "metric": "prompt_level_strict_acc,none",
+              "kind": "lm",
+              "weight": 0.33,
+              "quick_n": 150
+            }
+          ]
+        },
+        {
+          "key": "code",
+          "label": "Code",
+          "tasks": [
+            {
+              "name": "livecodebench_gen",
+              "metric": "pass@1",
+              "kind": "code",
+              "weight": 1.0,
+              "quick_n": 200,
+              "note": "Kontaminiert, und zwar ungleich: LiveCodeBench v6 enthält Aufgaben von Mai 2023 bis April 2025, ein neueres Release gibt es nicht. Die Trainings-Cutoffs der gemessenen Modelle liegen teils dahinter (Stand September 2026: Gemma 4 Januar 2025, Nemotron 3.5 Lightning September 2025, Qwen3.8 nicht veröffentlicht). Ein Modell mit späterem Cutoff kann einen größeren Teil der Aufgaben im Training gesehen haben; Abstände zwischen Modellen sind hier nur eingeschränkt aussagekräftig. Ausgaben werden exakt wie im offiziellen LiveCodeBench verglichen: 12 Aufgaben mit Fließkomma-Ausgabe sind dadurch für alle Modelle praktisch unlösbar (auch in den veröffentlichten LiveCodeBench-Zahlen); das senkt den erreichbaren Höchstwert für alle gleich. Liste und Abgleich mit dem offiziellen Auswerter im Kalibrierungsbericht (docs/superpowers/reports/2026-09-16-lcb-kalibrierung.md). Profil full lädt den vollen Datensatz und braucht dafür rund 9 GB RAM."
+            }
+          ]
+        },
+        {
+          "key": "german",
+          "label": "Deutsch",
+          "tasks": [
+            {
+              "name": "mmlu_prox_de",
+              "metric": "exact_match,custom-extract",
+              "kind": "lm",
+              "weight": 1.0,
+              "quick_n": 300
+            },
+            {
+              "name": "include_de_gen",
+              "metric": "exact_match,letter",
+              "kind": "lm",
+              "weight": 0.5,
+              "quick_n": 154,
+              "note": "Nur 154 Aufgaben (Test und Validation von INCLUDE-German, alle werden gemessen). Das 95-%-Intervall beträgt rund ±8 Punkte; kleinere Abstände zwischen Modellen sind Rauschen. Deshalb halbes Gewicht neben mmlu_prox_de."
+            }
+          ]
+        }
+      ]
+    },
     {
       "version": 2,
       "label": "v2",
@@ -84,11 +186,26 @@
     }
   ],
   "overview": [
+    {
+      "suite_version": 3,
+      "profile": "quick",
+      "runs": [
+        15,
+        21,
+        22,
+        20,
+        17,
+        16,
+        19
+      ]
+    },
     {
       "suite_version": 2,
       "profile": "quick",
       "runs": [
         8,
+        14,
+        12,
         10,
         7,
         6,
@@ -861,6 +978,1502 @@
           "estimated": false
         }
       ]
+    },
+    {
+      "id": 12,
+      "model": "gemma-4-12b-it",
+      "quant": "Q8",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-17",
+      "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": 82.04597276445436,
+      "categories": {
+        "knowledge": 71.23232323232322,
+        "math": 91.82222222222224,
+        "code": 86.79601226993864,
+        "german": 78.33333333333333
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 79.33333333333333,
+          "stderr": 2.341679610131769,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 63.13131313131313,
+          "stderr": 3.4373055019806156,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 92.4,
+          "stderr": 1.6793573067859682,
+          "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": 95.0920245398773,
+          "stderr": 1.6973278252200055,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 78.5,
+          "stderr": 2.9122423970017453,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 78.33333333333333,
+          "stderr": 2.382504669967188,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 16.51787775562755,
+          "p10": 16.50838245037658,
+          "p90": 16.521081944993295,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 851.2285976316333,
+          "p10": 772.9179519933921,
+          "p90": 852.9303345017652,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 1359.0489942173106,
+          "p10": 1357.8739287948283,
+          "p90": 1445.2580228216675,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 1408.5374341646545,
+          "p10": 1406.8044229376228,
+          "p90": 1429.5952238880777,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 621.4546849951148,
+          "p10": 620.2151466626674,
+          "p90": 687.8989755176008,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 2952.064286917448,
+          "p10": 2782.652004947886,
+          "p90": 2954.619111167267,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 11338.711781892926,
+          "p10": 11172.374014696106,
+          "p90": 11352.684816718102,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 14,
+      "model": "gemma-4-12b-it-heretic",
+      "quant": "Q8",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 2,
+      "suite_label": "v2",
+      "started_at": "2026-09-17",
+      "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": 83.09445761293921,
+      "categories": {
+        "knowledge": 72.07070707070707,
+        "math": 93.51111111111112,
+        "code": 86.79601226993864,
+        "german": 80.0
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 80.0,
+          "stderr": 2.3132597278649607,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 64.14141414141415,
+          "stderr": 3.416903640391528,
+          "n_samples": 198
+        },
+        {
+          "task": "math_hard_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 93.60000000000001,
+          "stderr": 1.5510587134374116,
+          "n_samples": 250
+        },
+        {
+          "task": "ifeval",
+          "category": "math",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 93.33333333333333,
+          "stderr": 2.0435234445940464,
+          "n_samples": 150
+        },
+        {
+          "task": "humaneval_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 95.0920245398773,
+          "stderr": 1.6973278252200055,
+          "n_samples": 163
+        },
+        {
+          "task": "mbpp_plus_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 78.5,
+          "stderr": 2.9122423970017453,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 80.0,
+          "stderr": 2.3132597278649607,
+          "n_samples": 300
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 16.51271607105954,
+          "p10": 16.507039849683782,
+          "p90": 16.515248096304745,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 853.9742669340698,
+          "p10": 757.1143617488933,
+          "p90": 858.8018268763251,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 1358.107397713914,
+          "p10": 1353.9826149807318,
+          "p90": 1447.551668426772,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 1386.1391956815926,
+          "p10": 1383.223042561945,
+          "p90": 1422.1981024185563,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 619.4566048216075,
+          "p10": 615.9744923003018,
+          "p90": 705.0649913493544,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 2954.110997961834,
+          "p10": 2778.05138588883,
+          "p90": 2963.1126631982625,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 11521.930878050625,
+          "p10": 11231.013681180775,
+          "p90": 11546.221878007054,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 15,
+      "model": "qwen3.8-flash-next-ple4",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-17",
+      "finished_at": "2026-09-18",
+      "adopted_from_date": "2026-09-15",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 71.45393635604162,
+      "categories": {
+        "knowledge": 77.86868686868686,
+        "math": 73.2,
+        "instructions": 49.56390977443609,
+        "code": 75.5,
+        "german": 81.13708513708514
+      },
+      "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": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 70.0,
+          "stderr": 2.904089347757578,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 73.2,
+          "stderr": 2.8068762382526717,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 36.0,
+          "stderr": 2.775911673437953,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 90.66666666666666,
+          "stderr": 2.383137380283291,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 75.5,
+          "stderr": 3.048807329211419,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 85.66666666666667,
+          "stderr": 2.0264875529605346,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 72.07792207792207,
+          "stderr": 3.6268473829697125,
+          "n_samples": 154
+        }
+      ],
+      "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": 16,
+      "model": "nemotron-3.5-lightning-dspark",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-17",
+      "finished_at": "2026-09-18",
+      "adopted_from_date": "2026-09-15",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 53.13354497354497,
+      "categories": {
+        "knowledge": 55.756902356902366,
+        "math": 40.400000000000006,
+        "instructions": 46.0,
+        "code": 53.0,
+        "german": 70.5108225108225
+      },
+      "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": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 39.2,
+          "stderr": 3.093820762040122,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 40.400000000000006,
+          "stderr": 3.1096688184825356,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 35.0,
+          "stderr": 2.758386421836852,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 79.33333333333333,
+          "stderr": 3.3171868444173502,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 53.0,
+          "stderr": 3.5380203419000424,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 72.0,
+          "stderr": 2.5966276044877823,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 67.53246753246754,
+          "stderr": 3.785604468377516,
+          "n_samples": 154
+        }
+      ],
+      "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": 17,
+      "model": "gemma-4-12b-it-heretic",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-18",
+      "finished_at": "2026-09-18",
+      "adopted_from_date": "2026-09-15",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 57.7166917293233,
+      "categories": {
+        "knowledge": 63.70909090909091,
+        "math": 48.4,
+        "instructions": 46.81203007518797,
+        "code": 62.0,
+        "german": 67.66233766233766
+      },
+      "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": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 58.4,
+          "stderr": 3.1235856237014503,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 48.4,
+          "stderr": 3.1669985030107437,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 33.0,
+          "stderr": 2.7193103591593655,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 88.66666666666667,
+          "stderr": 2.5969637523072806,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 62.0,
+          "stderr": 3.440813273035823,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 70.0,
+          "stderr": 2.6501719512585327,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 62.98701298701299,
+          "stderr": 3.903524043561299,
+          "n_samples": 154
+        }
+      ],
+      "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": 19,
+      "model": "qwen3.5-9b-heretic-v2-mtp",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-18",
+      "finished_at": "2026-09-18",
+      "adopted_from_date": "2026-09-16",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 51.662634870003295,
+      "categories": {
+        "knowledge": 65.07138047138046,
+        "math": 35.6,
+        "instructions": 44.140350877192986,
+        "code": 46.5,
+        "german": 67.00144300144301
+      },
+      "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": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 60.8,
+          "stderr": 3.093820762040122,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 35.6,
+          "stderr": 3.0343680657153196,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 29.666666666666668,
+          "stderr": 2.6416749751055386,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 88.0,
+          "stderr": 2.6621886338401435,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 46.5,
+          "stderr": 3.5357115664894154,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 69.33333333333334,
+          "stderr": 2.6666666666666665,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 62.33766233766234,
+          "stderr": 3.917267027864408,
+          "n_samples": 154
+        }
+      ],
+      "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
+        }
+      ]
+    },
+    {
+      "id": 20,
+      "model": "qwen3.8-27b-mtp",
+      "quant": "NVFP4",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-18",
+      "finished_at": "2026-09-18",
+      "adopted_from_date": "2026-09-16",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 64.20301055669476,
+      "categories": {
+        "knowledge": 69.85656565656565,
+        "math": 60.4,
+        "instructions": 49.498746867167924,
+        "code": 65.0,
+        "german": 76.25974025974027
+      },
+      "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": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 65.60000000000001,
+          "stderr": 3.0104503392316437,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 60.4,
+          "stderr": 3.09931978545779,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 38.333333333333336,
+          "stderr": 2.811757974289912,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 83.33333333333334,
+          "stderr": 3.0530971060304655,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_gen",
+          "category": "code",
+          "metric": "pass@1",
+          "score": 65.0,
+          "stderr": 3.3811478530256704,
+          "n_samples": 200
+        },
+        {
+          "task": "mmlu_prox_de",
+          "category": "german",
+          "metric": "exact_match,custom-extract",
+          "score": 79.0,
+          "stderr": 2.355524354210245,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 70.77922077922078,
+          "stderr": 3.676656608352373,
+          "n_samples": 154
+        }
+      ],
+      "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": 21,
+      "model": "gemma-4-12b-it",
+      "quant": "Q8",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-18",
+      "finished_at": "2026-09-19",
+      "adopted_from_date": "2026-09-17",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 65.82462114883167,
+      "categories": {
+        "knowledge": 67.48821548821549,
+        "math": 60.8,
+        "instructions": 51.318295739348365,
+        "code": 75.0,
+        "german": 74.51659451659451
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 79.33333333333333,
+          "stderr": 2.341679610131769,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 63.13131313131313,
+          "stderr": 3.4373055019806156,
+          "n_samples": 198
+        },
+        {
+          "task": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 60.0,
+          "stderr": 3.1046021028253312,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 60.8,
+          "stderr": 3.093820762040122,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 38.333333333333336,
+          "stderr": 2.811757974289912,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 90.66666666666666,
+          "stderr": 2.383137380283291,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_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": 78.33333333333333,
+          "stderr": 2.382504669967188,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 66.88311688311688,
+          "stderr": 3.8048475825209844,
+          "n_samples": 154
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 16.51787775562755,
+          "p10": 16.50838245037658,
+          "p90": 16.521081944993295,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 851.2285976316333,
+          "p10": 772.9179519933921,
+          "p90": 852.9303345017652,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 1359.0489942173106,
+          "p10": 1357.8739287948283,
+          "p90": 1445.2580228216675,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 1408.5374341646545,
+          "p10": 1406.8044229376228,
+          "p90": 1429.5952238880777,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 621.4546849951148,
+          "p10": 620.2151466626674,
+          "p90": 687.8989755176008,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 2952.064286917448,
+          "p10": 2782.652004947886,
+          "p90": 2954.619111167267,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 11338.711781892926,
+          "p10": 11172.374014696106,
+          "p90": 11352.684816718102,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
+    },
+    {
+      "id": 22,
+      "model": "gemma-4-12b-it-heretic",
+      "quant": "Q8",
+      "reasoning": "off",
+      "profile": "quick",
+      "suite_version": 3,
+      "suite_label": "v2b",
+      "started_at": "2026-09-19",
+      "finished_at": "2026-09-19",
+      "adopted_from_date": "2026-09-17",
+      "hardware": "NVIDIA DGX Spark, 128 GB RAM",
+      "provider": "vLLM 0.26.0",
+      "route": "direct",
+      "speed_comparable": true,
+      "overall": 65.82311637680058,
+      "categories": {
+        "knowledge": 68.44713804713804,
+        "math": 59.199999999999996,
+        "instructions": 49.97493734335839,
+        "code": 75.0,
+        "german": 76.4935064935065
+      },
+      "tasks": [
+        {
+          "task": "mmlu_pro",
+          "category": "knowledge",
+          "metric": "exact_match,custom-extract",
+          "score": 80.0,
+          "stderr": 2.3132597278649607,
+          "n_samples": 300
+        },
+        {
+          "task": "gpqa_diamond_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 64.14141414141415,
+          "stderr": 3.416903640391528,
+          "n_samples": 198
+        },
+        {
+          "task": "musr_gen",
+          "category": "knowledge",
+          "metric": "exact_match,letter",
+          "score": 61.199999999999996,
+          "stderr": 3.0881038748993976,
+          "n_samples": 250
+        },
+        {
+          "task": "omni_math_gen",
+          "category": "math",
+          "metric": "exact_match,boxed",
+          "score": 59.199999999999996,
+          "stderr": 3.1145209846548507,
+          "n_samples": 250
+        },
+        {
+          "task": "ifbench",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 35.66666666666667,
+          "stderr": 2.7702163901059955,
+          "n_samples": 300
+        },
+        {
+          "task": "ifeval",
+          "category": "instructions",
+          "metric": "prompt_level_strict_acc,none",
+          "score": 93.33333333333333,
+          "stderr": 2.0435234445940464,
+          "n_samples": 150
+        },
+        {
+          "task": "livecodebench_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": 80.0,
+          "stderr": 2.3132597278649607,
+          "n_samples": 300
+        },
+        {
+          "task": "include_de_gen",
+          "category": "german",
+          "metric": "exact_match,letter",
+          "score": 69.48051948051948,
+          "stderr": 3.7228400859666806,
+          "n_samples": 154
+        }
+      ],
+      "speed": [
+        {
+          "metric": "gen_tps",
+          "target_len": 0,
+          "context_len": null,
+          "value": 16.51271607105954,
+          "p10": 16.507039849683782,
+          "p90": 16.515248096304745,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 853.9742669340698,
+          "p10": 757.1143617488933,
+          "p90": 858.8018268763251,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 1358.107397713914,
+          "p10": 1353.9826149807318,
+          "p90": 1447.551668426772,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "pp_tps",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 1386.1391956815926,
+          "p10": 1383.223042561945,
+          "p90": 1422.1981024185563,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 512,
+          "context_len": 529,
+          "value": 619.4566048216075,
+          "p10": 615.9744923003018,
+          "p90": 705.0649913493544,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 4096,
+          "context_len": 4012,
+          "value": 2954.110997961834,
+          "p10": 2778.05138588883,
+          "p90": 2963.1126631982625,
+          "repeats": 5,
+          "estimated": false
+        },
+        {
+          "metric": "ttft_ms",
+          "target_len": 16384,
+          "context_len": 15971,
+          "value": 11521.930878050625,
+          "p10": 11231.013681180775,
+          "p90": 11546.221878007054,
+          "repeats": 5,
+          "estimated": false
+        }
+      ]
     }
   ]
 }

runs/10.html

@@ -174,7 +174,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/11.html

@@ -174,7 +174,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/12.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 – 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</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>Q8</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-17</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">82.0</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">71.2</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">91.8</td></tr>
+<tr><td>Code</td><td class="num">86.8</td></tr>
+<tr><td>Deutsch</td><td class="num">78.3</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">79.3</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">63.1</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">92.4</td>
+  <td class="num">± 1.7</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">95.1</td>
+  <td class="num">± 1.7</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">78.5</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">78.3</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">16.5</td>
+  <td class="num">16.5</td>
+  <td class="num">16.5</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">851.2</td>
+  <td class="num">772.9</td>
+  <td class="num">852.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">1359.0</td>
+  <td class="num">1357.9</td>
+  <td class="num">1445.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">1408.5</td>
+  <td class="num">1406.8</td>
+  <td class="num">1429.6</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">621.5</td>
+  <td class="num">620.2</td>
+  <td class="num">687.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">2952.1</td>
+  <td class="num">2782.7</td>
+  <td class="num">2954.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">11338.7</td>
+  <td class="num">11172.4</td>
+  <td class="num">11352.7</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/14.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>Q8</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2</dd>
+  <dt>Datum</dt><dd>2026-09-17</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">83.1</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">72.1</td></tr>
+<tr><td>Mathe &amp; Instruktionen</td><td class="num">93.5</td></tr>
+<tr><td>Code</td><td class="num">86.8</td></tr>
+<tr><td>Deutsch</td><td class="num">80.0</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">80.0</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">64.1</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">93.6</td>
+  <td class="num">± 1.6</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe &amp; Instruktionen</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>humaneval_plus_gen</td>
+  <td class="num">95.1</td>
+  <td class="num">± 1.7</td>
+  <td class="num">163</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>mbpp_plus_gen</td>
+  <td class="num">78.5</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">80.0</td>
+  <td class="num">± 2.3</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">16.5</td>
+  <td class="num">16.5</td>
+  <td class="num">16.5</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">854.0</td>
+  <td class="num">757.1</td>
+  <td class="num">858.8</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">1358.1</td>
+  <td class="num">1354.0</td>
+  <td class="num">1447.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">1386.1</td>
+  <td class="num">1383.2</td>
+  <td class="num">1422.2</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">619.5</td>
+  <td class="num">616.0</td>
+  <td class="num">705.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">2954.1</td>
+  <td class="num">2778.1</td>
+  <td class="num">2963.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">11521.9</td>
+  <td class="num">11231.0</td>
+  <td class="num">11546.2</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/15.html

@@ -0,0 +1,196 @@
+<!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#v3-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>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-17 – 2026-09-18<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-15 (Suite-Wechsel)</div></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">71.5</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">77.9</td></tr>
+<tr><td>Mathe</td><td class="num">73.2</td></tr>
+<tr><td>Instruktionen</td><td class="num">49.6</td></tr>
+<tr><td>Code</td><td class="num">75.5</td></tr>
+<tr><td>Deutsch</td><td class="num">81.1</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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">70.0</td>
+  <td class="num">± 2.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">73.2</td>
+  <td class="num">± 2.8</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">36.0</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_gen</td>
+  <td class="num">75.5</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">85.7</td>
+  <td class="num">± 2.0</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">72.1</td>
+  <td class="num">± 3.6</td>
+  <td class="num">154</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 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/16.html

@@ -0,0 +1,196 @@
+<!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#v3-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>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-17 – 2026-09-18<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-15 (Suite-Wechsel)</div></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">53.1</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">55.8</td></tr>
+<tr><td>Mathe</td><td class="num">40.4</td></tr>
+<tr><td>Instruktionen</td><td class="num">46.0</td></tr>
+<tr><td>Code</td><td class="num">53.0</td></tr>
+<tr><td>Deutsch</td><td class="num">70.5</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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">39.2</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">40.4</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">35.0</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_gen</td>
+  <td class="num">53.0</td>
+  <td class="num">± 3.5</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>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">67.5</td>
+  <td class="num">± 3.8</td>
+  <td class="num">154</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 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/17.html

@@ -0,0 +1,196 @@
+<!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#v3-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>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-18<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-15 (Suite-Wechsel)</div></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">57.7</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">63.7</td></tr>
+<tr><td>Mathe</td><td class="num">48.4</td></tr>
+<tr><td>Instruktionen</td><td class="num">46.8</td></tr>
+<tr><td>Code</td><td class="num">62.0</td></tr>
+<tr><td>Deutsch</td><td class="num">67.7</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">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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">58.4</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">48.4</td>
+  <td class="num">± 3.2</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">33.0</td>
+  <td class="num">± 2.7</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_gen</td>
+  <td class="num">62.0</td>
+  <td class="num">± 3.4</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>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">63.0</td>
+  <td class="num">± 3.9</td>
+  <td class="num">154</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 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/19.html

@@ -0,0 +1,196 @@
+<!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#v3-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>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-18<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-16 (Suite-Wechsel)</div></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">51.7</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">65.1</td></tr>
+<tr><td>Mathe</td><td class="num">35.6</td></tr>
+<tr><td>Instruktionen</td><td class="num">44.1</td></tr>
+<tr><td>Code</td><td class="num">46.5</td></tr>
+<tr><td>Deutsch</td><td class="num">67.0</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">76.3</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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">60.8</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">35.6</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">29.7</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_gen</td>
+  <td class="num">46.5</td>
+  <td class="num">± 3.5</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>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">62.3</td>
+  <td class="num">± 3.9</td>
+  <td class="num">154</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 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/20.html

@@ -0,0 +1,196 @@
+<!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#v3-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>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-18<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-16 (Suite-Wechsel)</div></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">64.2</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">69.9</td></tr>
+<tr><td>Mathe</td><td class="num">60.4</td></tr>
+<tr><td>Instruktionen</td><td class="num">49.5</td></tr>
+<tr><td>Code</td><td class="num">65.0</td></tr>
+<tr><td>Deutsch</td><td class="num">76.3</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">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>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">65.6</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">60.4</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">38.3</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_gen</td>
+  <td class="num">65.0</td>
+  <td class="num">± 3.4</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>
+<tr>
+  <td>Deutsch</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">± 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 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/21.html

@@ -0,0 +1,196 @@
+<!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 – 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#v3-quick">← Übersicht</a></p>
+<h1>gemma-4-12b-it</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>Q8</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-18 – 2026-09-19<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-17 (Suite-Wechsel)</div></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">65.8</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">67.5</td></tr>
+<tr><td>Mathe</td><td class="num">60.8</td></tr>
+<tr><td>Instruktionen</td><td class="num">51.3</td></tr>
+<tr><td>Code</td><td class="num">75.0</td></tr>
+<tr><td>Deutsch</td><td class="num">74.5</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">79.3</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">63.1</td>
+  <td class="num">± 3.4</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">60.0</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">60.8</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">38.3</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>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>livecodebench_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">78.3</td>
+  <td class="num">± 2.4</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">66.9</td>
+  <td class="num">± 3.8</td>
+  <td class="num">154</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">16.5</td>
+  <td class="num">16.5</td>
+  <td class="num">16.5</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">851.2</td>
+  <td class="num">772.9</td>
+  <td class="num">852.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">1359.0</td>
+  <td class="num">1357.9</td>
+  <td class="num">1445.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">1408.5</td>
+  <td class="num">1406.8</td>
+  <td class="num">1429.6</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">621.5</td>
+  <td class="num">620.2</td>
+  <td class="num">687.9</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">2952.1</td>
+  <td class="num">2782.7</td>
+  <td class="num">2954.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">11338.7</td>
+  <td class="num">11172.4</td>
+  <td class="num">11352.7</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/22.html

@@ -0,0 +1,196 @@
+<!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#v3-quick">← Übersicht</a></p>
+<h1>gemma-4-12b-it-heretic</h1>
+<dl class="meta">
+  <dt>Quantisierung</dt><dd>Q8</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profil</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Datum</dt><dd>2026-09-19<div class="sub">Ergebnisse übernommen aus Messung vom 2026-09-17 (Suite-Wechsel)</div></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">65.8</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">68.4</td></tr>
+<tr><td>Mathe</td><td class="num">59.2</td></tr>
+<tr><td>Instruktionen</td><td class="num">50.0</td></tr>
+<tr><td>Code</td><td class="num">75.0</td></tr>
+<tr><td>Deutsch</td><td class="num">76.5</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">80.0</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">64.1</td>
+  <td class="num">± 3.4</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Wissen &amp; Reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">61.2</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Mathe</td>
+  <td>omni_math_gen</td>
+  <td class="num">59.2</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruktionen</td>
+  <td>ifbench</td>
+  <td class="num">35.7</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruktionen</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">75.0</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">80.0</td>
+  <td class="num">± 2.3</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Deutsch</td>
+  <td>include_de_gen</td>
+  <td class="num">69.5</td>
+  <td class="num">± 3.7</td>
+  <td class="num">154</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">16.5</td>
+  <td class="num">16.5</td>
+  <td class="num">16.5</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">854.0</td>
+  <td class="num">757.1</td>
+  <td class="num">858.8</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">1358.1</td>
+  <td class="num">1354.0</td>
+  <td class="num">1447.6</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt-Verarbeitung [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">1386.1</td>
+  <td class="num">1383.2</td>
+  <td class="num">1422.2</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">619.5</td>
+  <td class="num">616.0</td>
+  <td class="num">705.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4012</td>
+  <td class="num">2954.1</td>
+  <td class="num">2778.1</td>
+  <td class="num">2963.1</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Zeit bis erstes Token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">15971</td>
+  <td class="num">11521.9</td>
+  <td class="num">11231.0</td>
+  <td class="num">11546.2</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Gemessen direkt am Inferenz-Server.</p>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/6.html

@@ -174,7 +174,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/7.html

@@ -174,7 +174,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/8.html

@@ -174,7 +174,7 @@
 </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>
+<footer class="page-foot">Stand 2026-09-20 · Scores 0–100 · erzeugt mit <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>