<?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[【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya]]></title><description><![CDATA[<p dir="auto">针对Deepseek V4.1 Flash 1017 + GLM 5.3做的运行框架。</p>
<p dir="auto"><a href="https://github.com/mw00/project-maya" rel="nofollow ugc">https://github.com/mw00/project-maya</a></p>
<p dir="auto">刚刚看到，我也正在安装准备测试！</p>
<p dir="auto">希望看到论坛各位大神的实测表现！</p>
]]></description><link>https://lcz.me/topic/2455</link><generator>RSS for Node</generator><lastBuildDate>Sun, 25 Oct 2026 00:57:19 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2455.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 20 Oct 2026 03:08:08 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 22:18:31 GMT]]></title><description><![CDATA[<p dir="auto">我也在折腾 DeepSeek 1017，做了一个 Strata-DS 分支，当前部署和实测记录放在这里：<a href="https://lcz.me/topic/2452">双 RTX PRO 6000 + 48GB RAM 跑 DeepSeek 1017</a>。</p>
<p dir="auto">接下来准备和 Project Maya 对齐测试条件做一轮 benchmark，比较 prefill/decode、资源占用和工具调用，也看看能互相借鉴哪些优化。对照结果测完再更新，敬请期待！</p>
]]></description><link>https://lcz.me/post/27483</link><guid isPermaLink="true">https://lcz.me/post/27483</guid><dc:creator><![CDATA[lunarmoon26]]></dc:creator><pubDate>Sat, 24 Oct 2026 22:18:31 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 07:23:21 GMT]]></title><description><![CDATA[<p dir="auto">4张4080系列 2x4080s + 2x4080 用strata 最新的0.2.4 1M上下文 ISTA DASLab  的Iq2xs 本地实测<br />
实际输入	Prefill	Decode	首 token（含思考）<br />
32,732 token	703 tok/s	43.3 tok/s	46.6 秒<br />
131,050 token	1,065 tok/s	42.9 tok/s	123.4 秒<br />
单人使用可用 但是 想爽玩还是的QFN q80</p>
]]></description><link>https://lcz.me/post/27309</link><guid isPermaLink="true">https://lcz.me/post/27309</guid><dc:creator><![CDATA[muskelon]]></dc:creator><pubDate>Sat, 24 Oct 2026 07:23:21 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 01:36:04 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f525.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--fire" style="height:1.4em;width:auto;vertical-align:middle" title="🔥" alt="🔥" /> 本帖参与讨论人数已达 <strong>10+</strong>，入选<strong>热议</strong>，作者 <strong>+3 积分</strong>奖励，感谢分享！</p>
<blockquote>
<p dir="auto">奖励凭证：热议 +3 分 · 编号 h2455-3984-1（系统自动发放，显示稍有延迟）</p>
</blockquote>
]]></description><link>https://lcz.me/post/27257</link><guid isPermaLink="true">https://lcz.me/post/27257</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:36:04 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 01:34:15 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tiger-tc" aria-label="Profile: tiger-tc">@<bdi>tiger-tc</bdi></a> 我 3090+3080 20g， 128G ddr4 3600， 也能跑到 25t/s</p>
]]></description><link>https://lcz.me/post/27256</link><guid isPermaLink="true">https://lcz.me/post/27256</guid><dc:creator><![CDATA[johnnybegood]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:34:15 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 01:18:31 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/charly-cheng" aria-label="Profile: Charly-Cheng">@<bdi>Charly-Cheng</bdi></a> 2张3090</p>
]]></description><link>https://lcz.me/post/27251</link><guid isPermaLink="true">https://lcz.me/post/27251</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:18:31 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 01:09:20 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a><br />
我也是DDR4,怎麼相差這麼多！難道是顯卡的關係？ 我用的顯卡是RTX 5000 Ada.<br />
還是你用雙顯卡？</p>
]]></description><link>https://lcz.me/post/27230</link><guid isPermaLink="true">https://lcz.me/post/27230</guid><dc:creator><![CDATA[Charly Cheng]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:09:20 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 00:34:14 GMT]]></title><description><![CDATA[<p dir="auto">其实strata的思路真的很适合mac studio统一内存。<br />
比如m5 ultra的cpu真的强的没边，用它来配合gpu，简直天衣无缝。<br />
而且统一内存还不用降低速度</p>
]]></description><link>https://lcz.me/post/27227</link><guid isPermaLink="true">https://lcz.me/post/27227</guid><dc:creator><![CDATA[babyedi3]]></dc:creator><pubDate>Sat, 24 Oct 2026 00:34:14 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Sat, 24 Oct 2026 00:05:29 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/charly-cheng" aria-label="Profile: Charly-Cheng">@<bdi>Charly-Cheng</bdi></a> 我是用ddr4 2633</p>
]]></description><link>https://lcz.me/post/27217</link><guid isPermaLink="true">https://lcz.me/post/27217</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Sat, 24 Oct 2026 00:05:29 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Fri, 23 Oct 2026 23:49:07 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a></p>
<p dir="auto">你的數據還不錯，應該是DDR5，我跑GLM 5.3，decode 只有 16 tok/s, 預填充也大約1k tok, 用polystrata 部署，decode 沒漲，預填充只有大約350 tok/s. 我關注GLM 5.3 模型，主要是他的專注力稀疏技術，在社群上的測試，是目前開源模型，在256k 上下文，長鏈任務，不會飄移的模型。<br />
剛部署起來，打算找時間測試。<br />
目前跑QFN IQ4 128k 上下文，長鏈任務飄移問題，有點頭痛，QFN Q4 256k, 長鏈任務，會因為長上下文召回率降低的問題，LLM 比較無法專注在任務主要的任務上。還需要叮嚀。<br />
打算換GLM 5.3，測試。</p>
]]></description><link>https://lcz.me/post/27208</link><guid isPermaLink="true">https://lcz.me/post/27208</guid><dc:creator><![CDATA[Charly Cheng]]></dc:creator><pubDate>Fri, 23 Oct 2026 23:49:07 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Fri, 23 Oct 2026 20:15:00 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/tiger-tc" aria-label="Profile: tiger-tc">@<bdi>tiger-tc</bdi></a> 这么牛逼，恭喜了！</p>
]]></description><link>https://lcz.me/post/27189</link><guid isPermaLink="true">https://lcz.me/post/27189</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Fri, 23 Oct 2026 20:15:00 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Fri, 23 Oct 2026 00:57:52 GMT]]></title><description><![CDATA[<p dir="auto">提交的代码合并进官方版本了!</p>
]]></description><link>https://lcz.me/post/26852</link><guid isPermaLink="true">https://lcz.me/post/26852</guid><dc:creator><![CDATA[tiger tc]]></dc:creator><pubDate>Fri, 23 Oct 2026 00:57:52 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Thu, 22 Oct 2026 08:46:27 GMT]]></title><description><![CDATA[<p dir="auto">我在我的A40，48G显存+768G内存上，把DeepSeek-V4.1-Flash-1017的Q8量化，模型全放内存读，平均跑到了26T/S。已经有实用价值了，给项目提了PR，看能不能合并进去！专给大内存用户使用！<br />
<a href="https://github.com/mw00/project-maya/pull/368" rel="nofollow ugc">https://github.com/mw00/project-maya/pull/368</a><br />
<img src="https://upload.lcz.me/uploads/c6c37fdc-8bab-4da4-8b22-9c3d91582da2.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/26718</link><guid isPermaLink="true">https://lcz.me/post/26718</guid><dc:creator><![CDATA[tiger tc]]></dc:creator><pubDate>Thu, 22 Oct 2026 08:46:27 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Wed, 21 Oct 2026 13:31:46 GMT]]></title><description><![CDATA[<p dir="auto">Deepseek V4.1 Flash 1017 的表現有機會強過 QFN</p>
<p dir="auto">GLM 5.3有人測試過好像不令人驚艷, TPS的效率開銷大</p>
]]></description><link>https://lcz.me/post/26532</link><guid isPermaLink="true">https://lcz.me/post/26532</guid><dc:creator><![CDATA[kos or]]></dc:creator><pubDate>Wed, 21 Oct 2026 13:31:46 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Wed, 21 Oct 2026 01:26:30 GMT]]></title><description><![CDATA[<p dir="auto">直接买就行了，只要你不在意隐私，或者搞两张华为96G显卡</p>
]]></description><link>https://lcz.me/post/26345</link><guid isPermaLink="true">https://lcz.me/post/26345</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Wed, 21 Oct 2026 01:26:30 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Tue, 20 Oct 2026 23:43:58 GMT]]></title><description><![CDATA[<p dir="auto">看了下各种硬件下的速度，我的是跑不动的~</p>
]]></description><link>https://lcz.me/post/26323</link><guid isPermaLink="true">https://lcz.me/post/26323</guid><dc:creator><![CDATA[joker_chang]]></dc:creator><pubDate>Tue, 20 Oct 2026 23:43:58 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Tue, 20 Oct 2026 23:49:05 GMT]]></title><description><![CDATA[<p dir="auto">这个其实没啥尝试的必要，开销太大，完全不如用在线，需要隐私抗审查，27B和QFN足够了。<br />
另外从仓库来看，不能证明是strata自己出的，估计碰瓷。</p>
]]></description><link>https://lcz.me/post/26320</link><guid isPermaLink="true">https://lcz.me/post/26320</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Tue, 20 Oct 2026 23:49:05 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Tue, 20 Oct 2026 03:13:55 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a> 谁能跟Strata Qwen3.8 flash next比，我现在看到100decode以下都无感，哈哈哈哈哈</p>
]]></description><link>https://lcz.me/post/26143</link><guid isPermaLink="true">https://lcz.me/post/26143</guid><dc:creator><![CDATA[JumpEggKiller]]></dc:creator><pubDate>Tue, 20 Oct 2026 03:13:55 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Tue, 20 Oct 2026 03:12:52 GMT]]></title><description><![CDATA[<p dir="auto">一直都有追踪<br />
但是glm 5.3 flash 实在跑不动<br />
50 decode, 1k prefill  没什么想玩的心</p>
]]></description><link>https://lcz.me/post/26142</link><guid isPermaLink="true">https://lcz.me/post/26142</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Tue, 20 Oct 2026 03:12:52 GMT</pubDate></item><item><title><![CDATA[Reply to 【最新】Strata专为Deepseek V4.1 Flash 1017 + GLM 5.3做的新分支-Project-Maya on Tue, 20 Oct 2026 03:12:38 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/05eff415-b5c8-4db2-8b70-b99a4e9980d1.jpeg" alt="图片.jpeg" class="img-markdown  img-fluid" /><br />
这里最重要的不是能够加载一个几百 B 的模型，而是三个工程能力：</p>
<ol>
<li>Expert Placement / Expert Cache<br />
将高频激活的专家尽可能保留在显存，剩余专家由 CPU 和内存处理，并在需要时从 NVMe 加载。这可以减少显存不足带来的性能损失。</li>
<li>适配超大上下文<br />
GLM 路线使用 KV Streaming，将部分 KV Cache 放到系统内存；DeepSeek 路线则使用模型自身的压缩注意力缓存格式。它们不是同一套实现。</li>
<li>推测解码和请求服务<br />
GLM 支持原生 MTP 或 DFlash2 推测解码，最多 8 路并发请求；DeepSeek 支持自己的 DSpark drafter，但不能直接假定拥有与 GLM 相同的并发性能。<br />
此外，Maya 提供 OpenAI 兼容 API、Anthropic 兼容 API、Tool Calling、网页聊天和硬件监控。</li>
</ol>
]]></description><link>https://lcz.me/post/26141</link><guid isPermaLink="true">https://lcz.me/post/26141</guid><dc:creator><![CDATA[JumpEggKiller]]></dc:creator><pubDate>Tue, 20 Oct 2026 03:12:38 GMT</pubDate></item></channel></rss>