Ergebnisse aktualisiert (9 Läufe)

app.js

@@ -33,6 +33,27 @@
     });
   });
 
+  // Zeilen nach Reasoning-Modus filtern. Die Schalter stehen nur im HTML, wenn eine Tabelle mehrere Modi
+  // zeigt; ohne JavaScript bleiben alle Zeilen sichtbar.
+  document.querySelectorAll("[data-filter-table]").forEach(function (group) {
+    var table = document.getElementById(group.dataset.filterTable);
+    if (!table || !table.tBodies[0]) return;
+    var buttons = Array.prototype.slice.call(group.querySelectorAll("[data-reasoning-filter]"));
+    buttons.forEach(function (button) {
+      button.addEventListener("click", function () {
+        var wanted = button.dataset.reasoningFilter;
+        buttons.forEach(function (other) {
+          var picked = other === button;
+          other.classList.toggle("active", picked);
+          other.setAttribute("aria-pressed", picked ? "true" : "false");
+        });
+        Array.prototype.slice.call(table.tBodies[0].rows).forEach(function (row) {
+          row.hidden = wanted !== "all" && (row.dataset.reasoning || "default") !== wanted;
+        });
+      });
+    });
+  });
+
   // Läufe für den Vergleich auswählen (mindestens 2).
   var picks = document.querySelectorAll(".compare-pick");
   var button = document.getElementById("compare-button");

index.html

@@ -19,15 +19,36 @@
 <p class="lead">Local language models compared: knowledge, math, code and German, plus speed. Each model variant is shown with its most recent completed run. How the numbers are measured is described under <a href="methodology.html">Methodology</a>.</p>
 <section id="v3-quick">
 <h2>Suite v2b · profile quick</h2>
+<p class="segmented" role="group" aria-label="Filter by reasoning mode" data-filter-table="table-v3-quick">
+  <button type="button" class="active" aria-pressed="true" data-reasoning-filter="all">All runs</button>
+  <button type="button" aria-pressed="false" data-reasoning-filter="off">Reasoning off</button>
+  <button type="button" aria-pressed="false" data-reasoning-filter="on">Reasoning on</button>
+</p>
 <div class="table-wrap">
-<table class="sortable">
+<table class="sortable" id="table-v3-quick">
 <thead><tr>
   <th data-sort="text">Model</th>
   <th data-sort="num" class="num">Overall</th>
 <th data-sort="num" class="num">Knowledge &amp; reasoning</th><th data-sort="num" class="num">Math</th><th data-sort="num" class="num">Instruction following</th><th data-sort="num" class="num">Code</th><th data-sort="num" class="num">German</th><th data-sort="num" class="num">gen t/s</th><th data-sort="num" class="num">pp t/s 4k</th><th data-sort="num" class="num">TTFT 4k ms</th>  <th data-sort="text">Date</th>
 </tr></thead>
 <tbody>
-<tr>
+<tr data-reasoning="on">
+  <td data-value="gpt-oss-120b">
+    <a href="runs/18.html" class="strong">gpt-oss-120b</a>
+    <div class="sub">MXFP4 · Reasoning on · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="73.08253056884635">73.1</td>
+  <td class="num" data-value="68.52727272727272" title="mmlu_pro: 71.0 ± 2.6 (n=300)&#10;gpqa_diamond_gen: 68.2 ± 3.3 (n=198)&#10;musr_gen: 66.4 ± 3.0 (n=250)&#10;">68.5</td>
+  <td class="num" data-value="70.39999999999999" title="omni_math_gen: 70.4 ± 2.9 (n=250)&#10;">70.4</td>
+  <td class="num strong" data-value="66.1203007518797" title="ifbench: 60.0 ± 2.8 (n=300)&#10;ifeval: 84.7 ± 3.0 (n=150)&#10;">66.1</td>
+  <td class="num strong" data-value="83.0" title="livecodebench_gen: 83.0 ± 2.7 (n=200)&#10;">83.0</td>
+  <td class="num" data-value="77.36507936507935" title="mmlu_prox_de: 80.3 ± 2.3 (n=300)&#10;include_de_gen: 71.4 ± 3.7 (n=154)&#10;">77.4</td>
+  <td class="num" data-value="">–</td>
+  <td class="num" data-value="">–</td>
+  <td class="num" data-value="">–</td>
+  <td data-value="2026-09-18">2026-09-18</td>
+</tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -36,14 +57,14 @@
   <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" 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</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -51,7 +72,7 @@
   <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="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>
@@ -59,7 +80,7 @@
   <td class="num" data-value="2952.064286917448">2952</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -75,7 +96,7 @@
   <td class="num" data-value="2954.110997961834">2954</td>
   <td data-value="2026-09-19">2026-09-19</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -91,7 +112,23 @@
   <td class="num" data-value="1577.6915270835161">1578</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
+  <td data-value="qwen3.8-27b-abliterated">
+    <a href="runs/23.html" class="strong">qwen3.8-27b-abliterated</a>
+    <div class="sub">NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM</div>
+  </td>
+  <td class="num strong" data-value="61.428784587731954">61.4</td>
+  <td class="num" data-value="71.50235690235691" title="mmlu_pro: 81.0 ± 2.3 (n=300)&#10;gpqa_diamond_gen: 70.7 ± 3.2 (n=198)&#10;musr_gen: 62.8 ± 3.1 (n=250)&#10;">71.5</td>
+  <td class="num" data-value="57.99999999999999" title="omni_math_gen: 58.0 ± 3.1 (n=250)&#10;">58.0</td>
+  <td class="num" data-value="46.07518796992481" title="ifbench: 34.0 ± 2.7 (n=300)&#10;ifeval: 82.7 ± 3.1 (n=150)&#10;">46.1</td>
+  <td class="num" data-value="59.5" title="livecodebench_gen: 59.5 ± 3.5 (n=200)&#10;">59.5</td>
+  <td class="num" data-value="72.06637806637806" title="mmlu_prox_de: 74.3 ± 2.5 (n=300)&#10;include_de_gen: 67.5 ± 3.8 (n=154)&#10;">72.1</td>
+  <td class="num" data-value="15.903617023850842">16</td>
+  <td class="num" data-value="2472.0116600865786">2472</td>
+  <td class="num" data-value="1627.824036986567">1628</td>
+  <td data-value="2026-09-20">2026-09-20</td>
+</tr>
+<tr data-reasoning="off">
   <td data-value="gemma-4-12b-it-heretic">
     <a href="runs/17.html" class="strong">gemma-4-12b-it-heretic</a>
     <div class="sub">NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -107,7 +144,7 @@
   <td class="num" data-value="763.404164230451">763</td>
   <td data-value="2026-09-18">2026-09-18</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -123,7 +160,7 @@
   <td class="num" data-value="554.1085170116276">554</td>
   <td data-value="2026-09-17">2026-09-17</td>
 </tr>
-<tr>
+<tr data-reasoning="off">
   <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 · direct · NVIDIA DGX Spark, 128 GB RAM</div>
@@ -144,7 +181,7 @@
 </div>
 </section>
 <p class="hint">Scores run from 0 to 100; the best value per category is shown in bold. Hovering a category cell lists its tasks with standard error (±) and number of items (n). Speed: generation in tokens/s, prompt processing at 4k context in tokens/s, time to first token at 4k context in ms. Click a column header to sort.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="app.js"></script>
 </body>

methodology.html

@@ -117,7 +117,7 @@ on hardware and inference software.</p>
 </tbody>
 </table>
 </div>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="app.js"></script>
 </body>

runs/15.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/16.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/17.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/18.html

@@ -0,0 +1,133 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>gpt-oss-120b – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>gpt-oss-120b</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>MXFP4</dd>
+  <dt>Reasoning</dt><dd>on</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-09-18 – 2026-09-19</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">73.1</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">68.5</td></tr>
+<tr><td>Math</td><td class="num">70.4</td></tr>
+<tr><td>Instruction following</td><td class="num">66.1</td></tr>
+<tr><td>Code</td><td class="num">83.0</td></tr>
+<tr><td>German</td><td class="num">77.4</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">71.0</td>
+  <td class="num">± 2.6</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">68.2</td>
+  <td class="num">± 3.3</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">66.4</td>
+  <td class="num">± 3.0</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">70.4</td>
+  <td class="num">± 2.9</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">60.0</td>
+  <td class="num">± 2.8</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">84.7</td>
+  <td class="num">± 3.0</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">83.0</td>
+  <td class="num">± 2.7</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">80.3</td>
+  <td class="num">± 2.3</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">71.4</td>
+  <td class="num">± 3.7</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr><td colspan="7" class="muted">No speed measurement.</td></tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

runs/19.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/20.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/21.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/22.html

@@ -188,7 +188,7 @@
 </table>
 </div>
 <p class="hint">Measured directly at the inference server.</p>
-<footer class="page-foot">As of 2026-09-20 · scores 0–100 · built with <code>bench publish</code></footer>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
 </main>
 <script src="../app.js"></script>
 </body>

runs/23.html

@@ -0,0 +1,195 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>qwen3.8-27b-abliterated – Local LLM Benchmarks</title>
+<link rel="stylesheet" href="../style.css">
+</head>
+<body>
+<header class="top">
+  <a class="brand" href="../index.html">Local LLM Benchmarks</a>
+  <nav>
+    <a href="../index.html">Overview</a>
+    <a href="../methodology.html">Methodology</a>
+  </nav>
+</header>
+<main>
+<p class="sub"><a href="../index.html#v3-quick">← Overview</a></p>
+<h1>qwen3.8-27b-abliterated</h1>
+<dl class="meta">
+  <dt>Quantization</dt><dd>NVFP4</dd>
+  <dt>Reasoning</dt><dd>off</dd>
+  <dt>Profile</dt><dd>quick</dd>
+  <dt>Suite</dt><dd>v2b</dd>
+  <dt>Date</dt><dd>2026-09-20 – 2026-09-21</dd>
+  <dt>Hardware</dt><dd>NVIDIA DGX Spark, 128 GB RAM</dd>
+  <dt>Inference software</dt><dd>vLLM 0.26.0</dd>
+  <dt>Measured</dt><dd>directly at the inference server</dd>
+  <dt>Overall</dt><dd class="strong">61.4</dd>
+</dl>
+
+<h2>Categories</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th class="num">Score</th></tr></thead>
+<tbody>
+<tr><td>Knowledge &amp; reasoning</td><td class="num">71.5</td></tr>
+<tr><td>Math</td><td class="num">58.0</td></tr>
+<tr><td>Instruction following</td><td class="num">46.1</td></tr>
+<tr><td>Code</td><td class="num">59.5</td></tr>
+<tr><td>German</td><td class="num">72.1</td></tr>
+</tbody>
+</table>
+</div>
+
+<h2>Tasks</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Category</th><th>Task</th><th class="num">Score</th><th class="num">±</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>mmlu_pro</td>
+  <td class="num">81.0</td>
+  <td class="num">± 2.3</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>gpqa_diamond_gen</td>
+  <td class="num">70.7</td>
+  <td class="num">± 3.2</td>
+  <td class="num">198</td>
+</tr>
+<tr>
+  <td>Knowledge &amp; reasoning</td>
+  <td>musr_gen</td>
+  <td class="num">62.8</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Math</td>
+  <td>omni_math_gen</td>
+  <td class="num">58.0</td>
+  <td class="num">± 3.1</td>
+  <td class="num">250</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifbench</td>
+  <td class="num">34.0</td>
+  <td class="num">± 2.7</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>Instruction following</td>
+  <td>ifeval</td>
+  <td class="num">82.7</td>
+  <td class="num">± 3.1</td>
+  <td class="num">150</td>
+</tr>
+<tr>
+  <td>Code</td>
+  <td>livecodebench_gen</td>
+  <td class="num">59.5</td>
+  <td class="num">± 3.5</td>
+  <td class="num">200</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>mmlu_prox_de</td>
+  <td class="num">74.3</td>
+  <td class="num">± 2.5</td>
+  <td class="num">300</td>
+</tr>
+<tr>
+  <td>German</td>
+  <td>include_de_gen</td>
+  <td class="num">67.5</td>
+  <td class="num">± 3.8</td>
+  <td class="num">154</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.</p>
+
+<h2>Speed</h2>
+<div class="table-wrap">
+<table>
+<thead><tr><th>Measure</th><th class="num">Context (target)</th><th class="num">Context (actual)</th><th class="num">Median</th><th class="num">p10</th><th class="num">p90</th><th class="num">n</th></tr></thead>
+<tbody>
+<tr>
+  <td>Generation [t/s]</td>
+  <td class="num">–</td>
+  <td class="num">–</td>
+  <td class="num">15.9</td>
+  <td class="num">15.1</td>
+  <td class="num">16.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">1805.5</td>
+  <td class="num">1685.0</td>
+  <td class="num">1818.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">2472.0</td>
+  <td class="num">2175.2</td>
+  <td class="num">2483.2</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Prompt processing [t/s]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">1685.4</td>
+  <td class="num">1489.9</td>
+  <td class="num">1688.7</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">512</td>
+  <td class="num">529</td>
+  <td class="num">293.0</td>
+  <td class="num">291.0</td>
+  <td class="num">314.3</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">4096</td>
+  <td class="num">4024</td>
+  <td class="num">1627.8</td>
+  <td class="num">1620.5</td>
+  <td class="num">1872.0</td>
+  <td class="num">5</td>
+</tr>
+<tr>
+  <td>Time to first token [ms]</td>
+  <td class="num">16384</td>
+  <td class="num">16001</td>
+  <td class="num">9493.9</td>
+  <td class="num">9475.4</td>
+  <td class="num">10858.2</td>
+  <td class="num">5</td>
+</tr>
+</tbody>
+</table>
+</div>
+<p class="hint">Measured directly at the inference server.</p>
+<footer class="page-foot">As of 2026-09-21 · scores 0–100 · built with <code>bench publish</code></footer>
+</main>
+<script src="../app.js"></script>
+</body>
+</html>

style.css

@@ -70,11 +70,13 @@ h1 { font-size: 20px; margin: 8px 0 16px; overflow-wrap: anywhere; }
 h2 { font-size: 16px; margin: 24px 0 8px; }
 h3 { font-size: 13px; margin: 8px 0 4px; color: var(--ink-2); }
 .toolbar { display: flex; flex-wrap: wrap; gap: 12px; align-items: center; justify-content: space-between; margin: 12px 0; }
-.segmented a { display: inline-block; padding: 6px 12px; border: 1px solid var(--border); color: var(--ink-2); }
-.segmented a:hover { text-decoration: none; color: var(--ink); }
-.segmented a:first-child { border-radius: 6px 0 0 6px; }
-.segmented a:last-child { border-radius: 0 6px 6px 0; }
-.segmented a.active { background: var(--surface); color: var(--ink); font-weight: 600; }
+.segmented a, .segmented button { display: inline-block; padding: 6px 12px; border: 1px solid var(--border); color: var(--ink-2); }
+.segmented button { background: none; border-radius: 0; font: inherit; }
+.segmented a:hover, .segmented button:hover { text-decoration: none; color: var(--ink); }
+.segmented a:first-child, .segmented button:first-child { border-radius: 6px 0 0 6px; }
+.segmented a:last-child, .segmented button:last-child { border-radius: 0 6px 6px 0; }
+.segmented a.active, .segmented button.active { background: var(--surface); color: var(--ink); font-weight: 600; }
+[hidden] { display: none !important; }
 
 button { font: inherit; padding: 6px 12px; border-radius: 6px; border: 1px solid var(--border); background: var(--surface); color: var(--ink); cursor: pointer; }
 button:disabled { opacity: 0.5; cursor: default; }