
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>akiraxtwo super</title>
      <link>https://akiraxclaw.com/blog</link>
      <description>整理值得留下來的 AI 工具觀察、實作片段與延伸閱讀。面向中文開發者與內容創作者。</description>
      <language>zh-TW</language>
      <managingEditor>akiraxtwo.dev@gmail.com (Akira)</managingEditor>
      <webMaster>akiraxtwo.dev@gmail.com (Akira)</webMaster>
      <lastBuildDate>Tue, 05 May 2026 00:00:00 GMT</lastBuildDate>
      <atom:link href="https://akiraxclaw.com/tags/推論加速/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://akiraxclaw.com/blog/gemma4-mtp-drafters-3x-speedup</guid>
    <title>Gemma 4 推論加速 3 倍：Multi-Token Prediction Drafters 完整技術解析</title>
    <link>https://akiraxclaw.com/blog/gemma4-mtp-drafters-3x-speedup</link>
    <description>Google 於 2026-05-05 為 Gemma 4 全系列釋出 MTP drafter 模型，利用 speculative decoding 達到最高 3.1 倍推論加速，且輸出逐 token 與原模型完全一致。本文深入拆解 drafter 架構設計、跨硬體實測數字、與 DeepSeek MTP 的本質差異，以及四個框架的實際使用方式。</description>
    <pubDate>Tue, 05 May 2026 00:00:00 GMT</pubDate>
    <author>akiraxtwo.dev@gmail.com (Akira)</author>
    <category>Gemma</category><category>Google</category><category>LLM</category><category>local-llm</category><category>推論加速</category><category>open-source</category><category>vLLM</category><category>Speculative Decoding</category>
  </item>

    </channel>
  </rss>
