<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>MATRIX · Naive-N0.5-Flash</title>
  <link>https://derekwang2002.github.io/ai-intelligence/projects/naive-n0-5-flash/</link>
  <description>AI 模型、Agent、工程与基础设施的重要变化，附一手来源与建议等级。</description>
  <language>zh-CN</language>
  <lastBuildDate>Mon, 02 Nov 2026 16:56:00 GMT</lastBuildDate>
  <atom:link href="https://derekwang2002.github.io/ai-intelligence/feeds/projects/naive-n0-5-flash.xml" rel="self" type="application/rss+xml"/>
  <item>
    <title>NaiveAI 开源原生 1M 上下文的 309B 稀疏模型</title>
    <link>https://derekwang2002.github.io/ai-intelligence/events/ev-20260927-01/</link>
    <guid isPermaLink="false">ev-20260927-01-published</guid>
    <pubDate>Sat, 03 Oct 2026 16:01:57 GMT</pubDate>
    <category>新事件</category>
    <category>TRIAL</category>
    <description>&lt;p&gt;NaiveAI 以 MIT 许可发布 Naive-N0.5-Flash：总参数 309B、激活 15.5B，原生上下文 1M token。模型用滑动窗口与稀疏注意力组合替代全注意力层，并公开权重和推理代码。部署要求支持 FP8 的 NVIDIA 硬件，仅权重约占 315 GB；基准仍由作者报告。&lt;/p&gt;&lt;p&gt;建议等级：TRIAL&lt;/p&gt;&lt;p&gt;来源：&lt;a href=&quot;https://huggingface.co/NaiveAI/Naive-N0.5-Flash&quot;&gt;Naive-N0.5-Flash model card and weights&lt;/a&gt; · &lt;a href=&quot;https://naive.ai/en/research/&quot;&gt;NaiveAI technical research&lt;/a&gt;&lt;/p&gt;</description>
  </item>
</channel>
</rss>
