<?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[【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF]]></title><description><![CDATA[<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f31f.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--star2" style="height:23px;width:auto;vertical-align:middle" title="🌟" alt="🌟" /> 模型简介</h2>
<p dir="auto">首先申明Qwen3.8‑35B‑A3B‑Distill‑GGUF不是Qwen官方发布的模型， 是Empero将 Qwen3.8 前沿模型蒸馏到 Qwen3.6‑35B‑A3B 的 Mixture‑of‑Experts（MoE）架构版本，提供多种 GGUF 量化文件，可直接用于 llama.cpp、Ollama、LM Studio、Jan、KoboldCpp 等推理框架。</p>
<p dir="auto">模型特点：</p>
<ul>
<li><strong>35B 总参数</strong></li>
<li>每 token 仅激活约 <strong>3B 参数</strong>（MoE 稀疏）</li>
<li>推理速度显著快于同体量稠密模型</li>
</ul>
<hr />
<h2>🧠 架构细节</h2>
<ul>
<li>30 层 <strong>Gated DeltaNet</strong></li>
<li>10 层 <strong>全注意力层</strong></li>
<li>256 个专家，每 token 路由 8 个专家</li>
</ul>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/26a0.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--warning" style="height:23px;width:auto;vertical-align:middle" title="⚠" alt="⚠" />️ 使用前注意</h2>
<p dir="auto">加载此模型需要 <strong>支持 Qwen3.6 / Gated DeltaNet MoE 的新版 llama.cpp</strong>。<br />
旧版本会直接加载失败。</p>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f4e6.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--package" style="height:23px;width:auto;vertical-align:middle" title="📦" alt="📦" /> GGUF 文件列表（量化方式与大小）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>文件名</th>
<th>量化</th>
<th>大小</th>
<th>说明</th>
</tr>
</thead>
<tbody>
<tr>
<td>IQ2_M</td>
<td>12.558 GB</td>
<td>最小可用，16GB 显卡可跑</td>
</tr>
<tr>
<td>Q2_K</td>
<td>13.839 GB</td>
<td>2-bit K-quant，兼容性好</td>
</tr>
<tr>
<td>IQ3_M</td>
<td>16.340 GB</td>
<td>3-bit，高性价比</td>
</tr>
<tr>
<td>Q3_K_M</td>
<td>17.664 GB</td>
<td>常规 3-bit K-quant</td>
</tr>
<tr>
<td>IQ4_XS</td>
<td>19.628 GB</td>
<td>接近 Q4_K_M，体积更小</td>
</tr>
<tr>
<td><strong>Q4_K_M（推荐）</strong></td>
<td>21.713 GB</td>
<td>质量与体积最佳平衡</td>
</tr>
<tr>
<td>Q5_K_M</td>
<td>25.348 GB</td>
<td>更高质量</td>
</tr>
<tr>
<td>Q6_K</td>
<td>29.209 GB</td>
<td>接近无损</td>
</tr>
<tr>
<td>Q8_0</td>
<td>37.802 GB</td>
<td>最高质量量化</td>
</tr>
<tr>
<td>BF16</td>
<td>71.067 GB</td>
<td>全精度参考版</td>
</tr>
<tr>
<td>Vision Projector</td>
<td>0.899 GB</td>
<td>图像输入需要搭配此文件</td>
</tr>
</tbody>
</table>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f4bb.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--computer" style="height:23px;width:auto;vertical-align:middle" title="💻" alt="💻" /> VRAM / RAM 需求指南</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>量化</th>
<th>GPU VRAM</th>
<th>CPU RAM</th>
</tr>
</thead>
<tbody>
<tr>
<td>IQ2_M / Q2_K</td>
<td>16GB</td>
<td>16GB</td>
</tr>
<tr>
<td>IQ3_M / Q3_K_M</td>
<td>20–24GB</td>
<td>24GB</td>
</tr>
<tr>
<td>IQ4_XS / Q4_K_M</td>
<td>24GB</td>
<td>32GB</td>
</tr>
<tr>
<td>Q5_K_M / Q6_K</td>
<td>32GB</td>
<td>48GB</td>
</tr>
<tr>
<td>Q8_0</td>
<td>48GB</td>
<td>64GB</td>
</tr>
<tr>
<td>BF16</td>
<td>80GB+</td>
<td>96GB</td>
</tr>
</tbody>
</table>
<p dir="auto">MoE 架构让 CPU 推理比稠密模型更可行。</p>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f6e0.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--hammer_and_wrench" style="height:23px;width:auto;vertical-align:middle" title="🛠" alt="🛠" /> 推理示例（llama.cpp）</h2>
<h3>文本推理</h3>
<pre><code class="language-bash">llama-cli -m Qwen3.8-35B-A3B-Q4_K_M.gguf \
  --temp 0.6 --top-p 0.95 --top-k 20 \
  -n 16384 -cnv
</code></pre>
<p dir="auto">模型属于 <strong>推理模型（reasoning model）</strong>，回答开头会带结构化块，需要在最终输出中清理掉。</p>
<hr />
<h3>图像推理（Vision）</h3>
<pre><code class="language-bash">llama-mtmd-cli \
  -m Qwen3.8-35B-A3B-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.8-35B-A3B-F16.gguf \
  --image photo.jpg \
  -p "Describe this image."
</code></pre>
<p dir="auto">视觉模型来自 Qwen3.6‑35B‑A3B，蒸馏过程未评估视觉能力。</p>
<hr />
<h2>🧩 Ollama / LM Studio / Jan / KoboldCpp</h2>
<p dir="auto">直接加载 GGUF 文件即可，聊天模板已内置。<br />
推荐采样参数：temperature=0.6, top_p=0.95, top_k=20</p>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f510.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--closed_lock_with_key" style="height:23px;width:auto;vertical-align:middle" title="🔐" alt="🔐" /> 校验下载</h2>
<p dir="auto"><a href="https://huggingface.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF/tree/main" rel="nofollow ugc">下载地址：Qwen3.8-35B-A3B-Distill-GGUF</a></p>
<pre><code class="language-bash">sha256sum -c SHA256SUMS --ignore-missing
</code></pre>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f4dc.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--scroll" style="height:23px;width:auto;vertical-align:middle" title="📜" alt="📜" /> 许可</h2>
<p dir="auto">模型权重使用 Apache‑2.0 许可。</p>
<hr />
]]></description><link>https://lcz.me/topic/1826</link><generator>RSS for Node</generator><lastBuildDate>Fri, 25 Sep 2026 21:28:13 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1826.rss" rel="self" type="application/rss+xml"/><pubDate>Sun, 20 Sep 2026 01:57:28 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Tue, 22 Sep 2026 04:40:40 GMT]]></title><description><![CDATA[<p dir="auto">谢谢楼主,我下载了,试一下跑手写</p>
<p dir="auto">报表识别效果如果,我之前的SGLANG 35B跑报表,有些细节老是识别不到</p>
]]></description><link>https://lcz.me/post/19971</link><guid isPermaLink="true">https://lcz.me/post/19971</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Tue, 22 Sep 2026 04:40:40 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Mon, 21 Sep 2026 11:54:16 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/marsempire" aria-label="Profile: marsempire">@<bdi>marsempire</bdi></a> 我的测试中和qwen3.8-27b相比代码水平互有胜负，qwen3.8-27b写的游戏往往第一次都会卡在启动画面，Qwen3.8‑35B写的游戏都是一次启动成功。肯定比Qwen3.6‑35B和Ornith-35B的代码水平强的多。无审查版我还是用Qwen3.6‑35B。</p>
]]></description><link>https://lcz.me/post/19752</link><guid isPermaLink="true">https://lcz.me/post/19752</guid><dc:creator><![CDATA[perter]]></dc:creator><pubDate>Mon, 21 Sep 2026 11:54:16 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Mon, 21 Sep 2026 10:11:07 GMT]]></title><description><![CDATA[<p dir="auto">和qwen3.8-27b相比，这个是不是代码水平不如27b，更适合写文章？<br />
习惯用无审查版，等无审查版的……</p>
]]></description><link>https://lcz.me/post/19730</link><guid isPermaLink="true">https://lcz.me/post/19730</guid><dc:creator><![CDATA[marsempire]]></dc:creator><pubDate>Mon, 21 Sep 2026 10:11:07 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Sun, 20 Sep 2026 09:56:51 GMT]]></title><description><![CDATA[<p dir="auto">看来是刚出来的35B 新产品</p>
]]></description><link>https://lcz.me/post/19532</link><guid isPermaLink="true">https://lcz.me/post/19532</guid><dc:creator><![CDATA[PENG XU]]></dc:creator><pubDate>Sun, 20 Sep 2026 09:56:51 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Sun, 20 Sep 2026 08:29:10 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/koala" aria-label="Profile: koala">@<bdi>koala</bdi></a> 大约比Qwen3.8-27B快一倍，开MTP差不多70token/s</p>
]]></description><link>https://lcz.me/post/19522</link><guid isPermaLink="true">https://lcz.me/post/19522</guid><dc:creator><![CDATA[perter]]></dc:creator><pubDate>Sun, 20 Sep 2026 08:29:10 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Sun, 20 Sep 2026 07:32:35 GMT]]></title><description><![CDATA[<p dir="auto">楼主有测试跑的速度如何吗？</p>
]]></description><link>https://lcz.me/post/19508</link><guid isPermaLink="true">https://lcz.me/post/19508</guid><dc:creator><![CDATA[koala]]></dc:creator><pubDate>Sun, 20 Sep 2026 07:32:35 GMT</pubDate></item><item><title><![CDATA[Reply to 【最强 35B MoE 模型】Qwen3.8‑35B‑A3B‑Distill‑GGUF on Sun, 20 Sep 2026 02:03:22 GMT]]></title><description><![CDATA[<p dir="auto">35B-A3B 的"速度" 是最適合本地Agent使用, 希望它的品質能盡快達到可用的程度, 操作會非常流暢, 會類似 5090 跑 Qwen3.8-27B的體感</p>
]]></description><link>https://lcz.me/post/19415</link><guid isPermaLink="true">https://lcz.me/post/19415</guid><dc:creator><![CDATA[kos or]]></dc:creator><pubDate>Sun, 20 Sep 2026 02:03:22 GMT</pubDate></item></channel></rss>