<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>门控残差 on AI 实战派 · 从技术到赚钱</title><link>https://guijiagi.com/tags/%E9%97%A8%E6%8E%A7%E6%AE%8B%E5%B7%AE/</link><description>Recent content in 门控残差 on AI 实战派 · 从技术到赚钱</description><generator>Hugo</generator><language>zh-cn</language><copyright>本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权</copyright><lastBuildDate>Thu, 15 Oct 2026 15:06:00 +0800</lastBuildDate><atom:link href="https://guijiagi.com/tags/%E9%97%A8%E6%8E%A7%E6%AE%8B%E5%B7%AE/index.xml" rel="self" type="application/rss+xml"/><item><title>Qwen3.8 Flash Next 架构拆解：512 个专家、4 分支门控残差、510 亿 n-gram 嵌入</title><link>https://guijiagi.com/posts/2026-10-15-02-qwen38-flash-next-512-experts-gated-residual/</link><pubDate>Thu, 15 Oct 2026 15:06:00 +0800</pubDate><guid>https://guijiagi.com/posts/2026-10-15-02-qwen38-flash-next-512-experts-gated-residual/</guid><description>Qwen3.8-Flash-Next 用 48 层、512 专家、10 路由加 1 共享，配合 4 分支门控残差与 4B MTP 头，把稀疏 MoE 推到新密度。</description></item></channel></rss>