<?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>Kolibri on Local First AI</title><link>https://localfirstai.eu/tags/kolibri/</link><description>Recent content in Kolibri on Local First AI</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Mon, 05 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://localfirstai.eu/tags/kolibri/index.xml" rel="self" type="application/rss+xml"/><item><title>We Ported Aleph Alpha's Kolibri to a MacBook. Our Own Gate Said No.</title><link>https://localfirstai.eu/posts/2026-10-05-our-own-gate-said-no/</link><pubDate>Mon, 05 Oct 2026 00:00:00 +0000</pubDate><guid>https://localfirstai.eu/posts/2026-10-05-our-own-gate-said-no/</guid><description>Aleph Alpha released Kolibri, a 78-billion-parameter English–German mixture-of-experts model, on 3 October, when no local runtime supported it. We wrote an MLX port, a separate numpy reference to check it against, and a pre-registered gate that had to pass before any score counted. It failed. By a rule we wrote before the run, this post has no Kolibri scores. It has what failed and what we found when we took the failure apart, including a batched mlx-lm expert layer that returned wrong numbers in a probe on our M5 Max. It also has what we&amp;#39;d do differently.</description></item></channel></rss>