Add the shipped models and their measured numbers

index.html

@@ -91,6 +91,33 @@ put the achievable number below it. Larger batches raise arithmetic intensity an
 change the picture entirely — this models the single-stream case, which is what
 most people run locally. Bandwidth figures are vendor peak.
 </p>
+<h2 style="margin-top:34px">Models built against this limit</h2>
+<p class="note" style="border:none;padding-top:0">
+Bytes are the only lever at batch one, so these are what the lever is worth in
+practice. Accuracy retained is the mean of ARC-Easy, ARC-Challenge, HellaSwag and
+PIQA against the uncompressed model, scored on the same GPU and harness. The
+speech model has no such benchmark, so it is measured by word error rate instead.
+</p>
+
+<div class="tblwrap"><table>
+<thead><tr><th>Model</th><th>Base</th><th>Size</th><th>Smaller by</th><th>Accuracy retained</th></tr></thead>
+<tbody>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Nunya-14B">Exeaon1-Nunya-14B</a></td><td>Qwen3-14B</td><td>7.38 GB</td><td>3.73x</td><td>100.3%</td></tr>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Kese-30B-A3B">Exeaon1-Kese-30B-A3B</a></td><td>Qwen3-30B-A3B</td><td>16.27 GB</td><td>3.50x</td><td>98.1%</td></tr>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Nunya-8B">Exeaon1-Nunya-8B</a></td><td>Qwen3-8B</td><td>4.11 GB</td><td>3.80x</td><td>99.6%</td></tr>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Dzo-4B">Exeaon1-Dzo-4B</a></td><td>Qwen3-4B</td><td>2.04 GB</td><td>3.67x</td><td>98.1%</td></tr>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Dzo-0.6B">Exeaon1-Dzo-0.6B</a></td><td>Qwen3-0.6B</td><td>0.31 GB</td><td>3.50x</td><td>96.3%</td></tr>
+<tr><td><a href="https://huggingface.co/Exeaon/Exeaon1-Voice-0.8B">Exeaon1-Voice-0.8B</a></td><td>Whisper v3-turbo</td><td>0.43 GB</td><td>3.47x</td><td>WER unchanged</td></tr>
+</tbody>
+</table></div>
+
+<p class="note">
+<strong>Retention is not a free lunch, and neither is speed.</strong> On a large
+GPU these run slower than dense fp16 — a vendor tensor-core GEMM is heavily
+tuned and there is bandwidth to spare, so trading compute for memory loses. The
+win is on the other side of this page: fitting where the dense model does not.
+</p>
+
 <p class="note" style="border:none;padding-top:0">
 Compressed models: <a href="https://huggingface.co/Exeaon">huggingface.co/Exeaon</a> ·
 Runtime: <a href="https://github.com/ExeaonLM">github.com/ExeaonLM</a> ·