Intel / post

Prism ML Releases Ternary Bonsai 2 27B, 9x Smaller, 98% Retention

Source: x.com
Date: Published Sep 17, 2026

Context

PrismML's Ternary Bonsai 2 27B is a compressed version of Qwen3.8 27B that, according to PrismML's announcement, keeps 98.2% of the original's aggregate performance while being 9 times smaller than the full-precision model. It takes 5.9 GB, which means a 27-billion-parameter model can be loaded with a fraction of the memory the original needs. It is available under the Apache 2.0 license.

This release changes quality rather than size. Two months after the first Bonsai 27B, the footprint is still 5.9 GB, and PrismML says the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, reasoning and long-horizon . Agentic coding and long-horizon tool use are the claims to watch, because they involve many steps in which a compressed model's small errors can add up.

The 98.2% figure is PrismML's own measurement and an aggregate, so it can hide larger drops on individual tasks, and the announcement's text does not say which benchmarks make up that aggregate or how each one moved. Until independent tests break the result down by task, the supported reading is that the model fits in 5.9 GB with a small claimed loss against Qwen3.8 27B, not that it matches the full model on every kind of work.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.

Read the source x.com