<?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>MLX on Fabian G. Williams</title><link>https://www.fabswill.com/tags/mlx/</link><description>Recent content in MLX on Fabian G. Williams</description><image><title>Fabian G. Williams</title><url>https://www.fabswill.com/img/main/logo.jpg</url><link>https://www.fabswill.com/img/main/logo.jpg</link></image><generator>Hugo</generator><language>en-us</language><lastBuildDate>Thu, 20 Aug 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://www.fabswill.com/tags/mlx/index.xml" rel="self" type="application/rss+xml"/><item><title>Doug Was Right: I Swapped In The MoE, And The Concurrency Math Changed</title><link>https://www.fabswill.com/blog/doug-was-right-i-measured-the-moe/</link><pubDate>Thu, 20 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.fabswill.com/blog/doug-was-right-i-measured-the-moe/</guid><description>Last post I measured the concurrency ceiling on a 27B dense model and closed with a promise. Doug Ware, who builds applied AI systems, had handed me the one caveat I did not test: a dense model is the hard case, and a mixture-of-experts model with only about 3B active parameters per token should leave real headroom for a second slot to pay off. So I ran the exact same 36-load matrix on an MLX MoE, changed one variable, and let the numbers settle it. The dense model flatlines. The MoE keeps climbing. And at 8 concurrent agents the dense model makes you wait 32 seconds for a first token while the MoE answers in under 1.</description></item><item><title>You Asked, So I Measured: What Concurrency Actually Costs On One Local Model</title><link>https://www.fabswill.com/blog/you-asked-i-measured-the-concurrency-ceiling/</link><pubDate>Tue, 18 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.fabswill.com/blog/you-asked-i-measured-the-concurrency-ceiling/</guid><description>Reddit pushed back on my batching post with 3 sharp, testable claims: decode is memory-bandwidth-bound, prefill batches better than decode, and prompt length changes the whole story. So I built a measurement matrix, ran 36 barrier-synchronized loads against 1 local Qwen 3.8 on my Macbook Pro M3 Max, and let the numbers settle it. 2 of the 3 predictions held. 1 did not show up the way I expected, and I am keeping the correction in. Then a real agent handoff stalled on the exact wall these charts describe, and I show how a one-day-old runtime was already a first-class citizen in my receipts governance framework.</description></item><item><title>I Added a Second Local Agent This Week. Here Is the Receipt for Every Human Decision Behind It.</title><link>https://www.fabswill.com/blog/i-added-a-second-local-agent-and-kept-the-receipt/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.fabswill.com/blog/i-added-a-second-local-agent-and-kept-the-receipt/</guid><description>I trialed a second local coding agent, Hermes from Nous Research, on my own MacBook, pointed at the same local Qwen 3.8 model I already run. Installed additively so nothing already working could break, governed with manual approvals, then tested until it proved it behaves. Here is the trial, in tables and screenshots, plus a receipt for the human hours behind writing it up.</description></item><item><title>Two Agents, One Local Model: Do They Run in Parallel, or Take Turns? I Measured It.</title><link>https://www.fabswill.com/blog/two-agents-one-local-model-batching/</link><pubDate>Mon, 17 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.fabswill.com/blog/two-agents-one-local-model-batching/</guid><description>A reader asked what happens if I run two local coding agents against the same Qwen 3.8 model on one Mac at the same time. I thought I knew the answer. I was wrong. So I read the server source, wrote a barrier-synchronized load driver to remove the human-ordering bias, and measured it. Batching is real up to 32 wide, but it is not free, and three innocent-looking choices collapse it back to a single lane.</description></item><item><title>I Swapped My Local Coding Model Overnight From a Hotel. The Agent Graded the Upgrade Itself.</title><link>https://www.fabswill.com/blog/local-model-upgrade-qwen-38-agent-self-tested/</link><pubDate>Sat, 15 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.fabswill.com/blog/local-model-upgrade-qwen-38-agent-self-tested/</guid><description>Local models are how I run my community work, my volunteer projects, my side hustles, and my musings, for two reasons: cost and keeping client data away from the labs. A new version dropped, so I upgraded it overnight from a hotel, additively, without touching the old one. Then I handed the new model the reviews and let my OpenCode agent test its own upgrade. It stood up a throwaway server, probed itself, and found three config gaps quietly throttling it. Here is the journey, in tables.</description></item></channel></rss>