<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Qwen3.8 27B M5 Max]]></title><description><![CDATA[<p dir="auto">我下载了全精度权重用 oMLX 转换成了 oQ4e 量化，保留 MTP<br />
<a href="https://huggingface.co/fevaoctwh/Qwen3.8-27B-oQ4e-mtp" rel="nofollow ugc">https://huggingface.co/fevaoctwh/Qwen3.8-27B-oQ4e-mtp</a><br />
在 M5 Max平台上可以说是精度和速度的一个平衡吧，更高的量化，M5 Max 就受限于内存带宽，decode 速度完全上不去了</p>
<p dir="auto">Measured with the built-in oMLX benchmark (<code>/admin</code> dashboard), single-request mode:</p>
<ul>
<li><strong>Engine:</strong> Auto</li>
<li><strong>Context:</strong> Code (Python)</li>
<li><strong>Test pattern:</strong> <code>pp&lt;T&gt; / tg128</code> = prefill T tokens, then generate 128 tokens</li>
<li><strong>Hardware:</strong> Apple Silicon (M5 Max, 128 GB unified memory)</li>
</ul>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Test</th>
<th>TTFT (ms)</th>
<th style="text-align:right">TPOT (ms)</th>
<th style="text-align:right">Prefill (tok/s)</th>
<th style="text-align:right">Generation (tok/s)</th>
<th style="text-align:right">E2E (s)</th>
<th style="text-align:right">Throughput (tok/s)</th>
<th style="text-align:right">Peak Mem</th>
</tr>
</thead>
<tbody>
<tr>
<td>pp1024 / tg128</td>
<td>1161.5</td>
<td style="text-align:right">15.84</td>
<td style="text-align:right">881.6</td>
<td style="text-align:right">63.6</td>
<td style="text-align:right">3.184</td>
<td style="text-align:right">361.9</td>
<td style="text-align:right">16.72 GB</td>
</tr>
<tr>
<td>pp4096 / tg128</td>
<td>5155.9</td>
<td style="text-align:right">19.23</td>
<td style="text-align:right">794.4</td>
<td style="text-align:right">52.4</td>
<td style="text-align:right">7.614</td>
<td style="text-align:right">554.8</td>
<td style="text-align:right">18.19 GB</td>
</tr>
<tr>
<td>pp8192 / tg128</td>
<td>11856.7</td>
<td style="text-align:right">18.13</td>
<td style="text-align:right">690.9</td>
<td style="text-align:right">55.6</td>
<td style="text-align:right">14.176</td>
<td style="text-align:right">586.9</td>
<td style="text-align:right">19.24 GB</td>
</tr>
<tr>
<td>pp16384 / tg128</td>
<td>25982.8</td>
<td style="text-align:right">17.80</td>
<td style="text-align:right">630.6</td>
<td style="text-align:right">56.6</td>
<td style="text-align:right">28.260</td>
<td style="text-align:right">584.3</td>
<td style="text-align:right">20.78 GB</td>
</tr>
</tbody>
</table>
<p dir="auto">量化后的能力肯定是打折的，作为 Agent 本地对话模型和简单任务是完全够用了，我现在使用的是Qwen3.6-35B-A3B-oQ6e<br />
M 芯片目前的架构还是适合激活参数小的 MoE， 甜点尺寸应该是 100B 左右总参数量但是激活小于 10B 参数的模型</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Model</th>
<th>Type</th>
<th>Released</th>
<th>MMLU-Pro</th>
<th>LiveCodeBench v6</th>
<th>SWE-bench Verified</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Qwen3.8-27B-oQ4e-mtp (this model)</strong></td>
<td><strong>Quantized 4/5-bit MLX, 27B dense</strong></td>
<td><strong>Aug 2026</strong></td>
<td><strong>80.7%</strong> *</td>
<td>— (not run)</td>
<td>—</td>
</tr>
<tr>
<td>Qwen3.6-27B</td>
<td>Non-quantized, 27B dense</td>
<td>Apr 2026</td>
<td>86.2%</td>
<td>~83.9</td>
<td>77.2%</td>
</tr>
<tr>
<td>Qwen3.6-35B-A3B</td>
<td>Non-quantized, 35B MoE</td>
<td>Apr 2026</td>
<td>85.2%</td>
<td>80.4</td>
<td>73.4%</td>
</tr>
<tr>
<td>Gemma 4 31B</td>
<td>Non-quantized, 31B dense</td>
<td>Mar 2026</td>
<td>85.2%</td>
<td>80.0</td>
<td>52.0%</td>
</tr>
</tbody>
</table>
<p dir="auto">LiveCodeBench v6 还在跑 27/300…… 结果要明天了</p>
]]></description><link>https://lcz.me/topic/1138/qwen3.8-27b-m5-max</link><generator>RSS for Node</generator><lastBuildDate>Sun, 30 Aug 2026 15:26:41 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1138.rss" rel="self" type="application/rss+xml"/><pubDate>Sat, 15 Aug 2026 14:14:14 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to Qwen3.8 27B M5 Max on Sun, 16 Aug 2026 00:53:40 GMT]]></title><description><![CDATA[<p dir="auto">我找了一下, 我是 5月19日测试的, Qwen 3.6 27B oQ4.</p>
<blockquote>
<p dir="auto">今天oMLX发布了 0.3.9rc1, 支持了 native MTP, 我又测了一下加上了MTP机制的 27b oQ4, decode 明显提升, PP 基本不变.</p>
</blockquote>
<p dir="auto"><img src="https://upload.lcz.me/uploads/0948842d-9a88-4bc5-a041-2c523b7b2960.png" alt="ad072ce5-9cd4-438d-983f-7117ed48cfb9.png" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/12346</link><guid isPermaLink="true">https://lcz.me/post/12346</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Sun, 16 Aug 2026 00:53:40 GMT</pubDate></item><item><title><![CDATA[Reply to Qwen3.8 27B M5 Max on Sun, 16 Aug 2026 00:50:16 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/fang-liu" aria-label="Profile: Fang-Liu">@<bdi>Fang-Liu</bdi></a></p>
<p dir="auto">这个速度已经相当惊人了. 我的M5pro, oMlx, 4bit量化, 也就在 17-20出点儿头. 好像当时也是MTP的模型</p>
<p dir="auto">你这个50-60的decode速度, 已经完全够用了. prefill 大几百, 稍微慢些, 但是也肯定够用.</p>
<p dir="auto">比我想象中的好多了. <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=6e8e1372d58" class="not-responsive emoji emoji-android emoji--+1" style="height:23px;width:auto;vertical-align:middle" title=":+1:" alt="👍" /> <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=6e8e1372d58" class="not-responsive emoji emoji-android emoji--+1" style="height:23px;width:auto;vertical-align:middle" title=":+1:" alt="👍" /> <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=6e8e1372d58" class="not-responsive emoji emoji-android emoji--+1" style="height:23px;width:auto;vertical-align:middle" title=":+1:" alt="👍" /></p>
]]></description><link>https://lcz.me/post/12345</link><guid isPermaLink="true">https://lcz.me/post/12345</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Sun, 16 Aug 2026 00:50:16 GMT</pubDate></item><item><title><![CDATA[Reply to Qwen3.8 27B M5 Max on Sat, 15 Aug 2026 17:24:55 GMT]]></title><description><![CDATA[<p dir="auto">苹果平台还是老实用Moe吧。</p>
]]></description><link>https://lcz.me/post/12334</link><guid isPermaLink="true">https://lcz.me/post/12334</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sat, 15 Aug 2026 17:24:55 GMT</pubDate></item></channel></rss>