<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Weightless on Matt Suiche</title><link>https://www.msuiche.com/tags/weightless/</link><description>Recent content in Weightless on Matt Suiche</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 06 Sep 2026 00:00:00 +0200</lastBuildDate><atom:link href="https://www.msuiche.com/tags/weightless/index.xml" rel="self" type="application/rss+xml"/><item><title>Steering a Loop: Control Vectors Meet the Looped Transformer</title><link>https://www.msuiche.com/posts/abliterating-a-loop-control-vectors-meet-the-looped-transformer/</link><pubDate>Sun, 06 Sep 2026 00:00:00 +0200</pubDate><guid>https://www.msuiche.com/posts/abliterating-a-loop-control-vectors-meet-the-looped-transformer/</guid><description>&lt;p&gt;&lt;a href="https://huggingface.co/Nanbeige/Nanbeige4.2-3B" target="_blank" rel="noopener"&gt;Nanbeige4.2-3B&lt;/a&gt; is the
first open-weight &lt;strong&gt;Looped Transformer&lt;/strong&gt; anyone can download: a 22-layer
stack executed twice, 44 effective layer-passes, 3B non-embedding
parameters, and benchmark numbers that embarrass dense models four times
its size. It buys that capacity with repetition rather than parameters.&lt;/p&gt;
&lt;p&gt;The lineage runs from the Universal Transformer through
&lt;a href="https://huggingface.co/tomg-group-umd/huginn-0125" target="_blank" rel="noopener"&gt;Huginn&lt;/a&gt; (a recurrent
block looped a variable number of times at inference) and ByteDance Seed&amp;rsquo;s
&lt;a href="https://huggingface.co/ByteDance/Ouro-2.6B" target="_blank" rel="noopener"&gt;Ouro&lt;/a&gt; (looped computation
trained into pretraining itself, with an entropy-regularized objective
that learns how many iterations an input deserves), to Google&amp;rsquo;s
&lt;a href="https://arxiv.org/abs/2507.10524" target="_blank" rel="noopener"&gt;Mixture-of-Recursions&lt;/a&gt;, which routes
each token to its own recursion depth. On September 1st, The Information
&lt;a href="https://tosea.ai/blog/looped-transformer-recurrent-depth-astra-guide" target="_blank" rel="noopener"&gt;reported&lt;/a&gt;
that OpenAI&amp;rsquo;s Astra is recurrent-depth: latent iteration in place of
visible chain-of-thought, which is why the safety community spent the
first week of September arguing about monitorability. OpenAI has not
published architecture details, so treat the report as a strong rumor.
Either way, depth has become a runtime quantity; Nanbeige4.5 is already
training with the token-routed machinery in its config.&lt;/p&gt;</description></item><item><title>Inkling on Two DGX Sparks: The Only vLLM Lane, and the Two Walls Behind It</title><link>https://www.msuiche.com/posts/inkling-on-two-dgx-sparks-the-only-vllm-lane-and-the-two-walls-behind-it/</link><pubDate>Fri, 04 Sep 2026 23:00:00 +0200</pubDate><guid>https://www.msuiche.com/posts/inkling-on-two-dgx-sparks-the-only-vllm-lane-and-the-two-walls-behind-it/</guid><description>&lt;p&gt;Thinking Machines shipped &lt;a href="https://thinkingmachines.ai/news/inkling-small/" target="_blank" rel="noopener"&gt;Inkling-Small&lt;/a&gt; at the end of July: a ~300B-total,
~10B-active hybrid MoE (short-conv plus relative-bias attention), natively
multimodal, 1M context, and - the part that matters for homelab hardware -
released as NVFP4 from day one. 170.7 GB of weights. Two DGX Sparks hold 243 GB
of unified memory. You can see where this is going.&lt;/p&gt;
&lt;p&gt;For five weeks, though, every Spark recipe for it -
&lt;a href="https://github.com/drowzeys/keys-1M-CTX-Inkling-Small-NVFP4-Dspark-NVFP4-KV-Cache-SGlang-SM121-optimized-on-Two-DGX-Sparks" target="_blank" rel="noopener"&gt;drowzeys&amp;rsquo; champion image&lt;/a&gt;,
&lt;a href="https://github.com/MiaAI-Lab/Inkling-Small-NVFP4-Dual-DGX-Sparks" target="_blank" rel="noopener"&gt;MiaAI&amp;rsquo;s wrapper&lt;/a&gt;,
and the half-dozen forks - ran on &lt;strong&gt;SGLang&lt;/strong&gt; with a custom-baked image, for a
simple reason: SGLang had Inkling support in early August, and vLLM only gained
it this week, in v0.28.0. The engine we build everything on could not even load
the model before Monday.&lt;/p&gt;</description></item></channel></rss>