<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Series on 67AI Lab</title>
    <link>https://67ailab.com/series/</link>
    <description>Recent content in Series on 67AI Lab</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Thu, 12 Nov 2026 15:10:00 +0000</lastBuildDate>
    <atom:link href="https://67ailab.com/series/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Small Models on One DGX Spark</title>
      <link>https://67ailab.com/series/small-models-dgx-spark/</link>
      <pubDate>Thu, 12 Nov 2026 15:10:00 +0000</pubDate>
      <guid>https://67ailab.com/series/small-models-dgx-spark/</guid>
      <description>&lt;p&gt;Every team using a frontier model eventually asks the same thing: &lt;em&gt;how much of&#xA;this could a small open model do on our own hardware?&lt;/em&gt; Benchmarks don&amp;rsquo;t&#xA;answer it, because leaderboard tasks aren&amp;rsquo;t your tasks, and a model that&amp;rsquo;s&#xA;&amp;ldquo;close on average&amp;rdquo; can fail in exactly the way that hurts you.&lt;/p&gt;&#xA;&lt;p&gt;So we ran the experiment. We took &lt;strong&gt;five open-weight models under 100B&#xA;parameters&lt;/strong&gt;, served them one at a time on a single&#xA;&lt;a href=&#34;https://www.nvidia.com/en-us/products/workstations/dgx-spark/&#34;&gt;NVIDIA DGX Spark&lt;/a&gt;&#xA;(GB10, 121 GB unified memory), and gave them &lt;strong&gt;eight narrow, realistic&#xA;tasks&lt;/strong&gt;. Each task had a pass rule written down before any model ran. Every&#xA;result was compared with &lt;strong&gt;Claude Opus 5.5&lt;/strong&gt; and &lt;strong&gt;DeepSeek-V4.1-Flash&lt;/strong&gt;, and&#xA;every surprising score was checked by reading the traces.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
