A full-size open model now fits in 5.9 gigabytes.
PrismML released Bonsai 2 on 17 September, a squeezed copy of Alibaba's Qwen3.8 27B model. Every weight is stored as minus one, zero or plus one, shrinking the file ninefold. PrismML says it keeps 98.2% of the original's benchmark score, but nobody has rechecked that.
Why it mattersA model this size used to need a server, and it now runs on a gaming card. The weights carry an Apache licence, so you can download and run them freely.
▶ Related video: Is Bonsai 2 27B REALLY 98% of Qwen 3.8 27B fp16?
Read in the app — free, in 9 languages
Related stories
Mistral gave away its trillion-parameter model, with one catch.2026-10-29Nvidia released an open model that writes down speech in 40 languages.
2026-10-29Google's Gemini 4 Carbon edged past Claude on a hard AI ranking.
2026-10-29Google's newest world model builds worlds you can hear and talk to.
2026-10-28OpenAI's new mid-size model costs half as much as the one before.
2026-10-27