A Hobbyist Shrank a Big AI in Half and It Still Works

A lone hobbyist developer has done something usually reserved for big labs: taken a large open AI model and squeezed it down to less than half its size without losing much of its ability. The original, Muse-Glimmer, has 30 billion parameters — think of parameters as the model's internal wiring. Instead of deleting whole sections, the developer trimmed the width of each layer, like slimming every floor of a building rather than knocking off the top ten.

The smaller model, Xyntetik-Kvist-14B, weighs in at about 14.4 billion parameters. That matters because it now fits on a single 24-gigabyte graphics card — the kind of hardware a serious hobbyist or small business might own — at 15.4 gigabytes compressed, or just 10.3 gigabytes in a tighter format. No data center, no monthly cloud bill. It handles the practical stuff: looking up contacts, weather, flights, currency conversions, dates, units and stock prices. It completes 57 of 60 test tasks, where its much bigger parent does 60. It's released under an Apache-2.0 license, which means anyone can use it commercially for free.

There are real catches. It's bad at arithmetic: give it a calculator tool and it's fine, but on its own it gets "17% of 2,340" wrong — and the developer says so plainly on the model card. Roughly 7 out of every 160 runs get stuck in a reasoning loop, spinning without finishing. It's also brand new, made by one person, and hasn't been widely tested by others yet.

What may be most unusual is the honesty. The developer published all 12 training attempts, including the failures, along with the defects and mid-course corrections. For anyone deciding whether to build on it, that's a rare look at exactly where the model breaks before they spend time on it.