PrismML shrinks Qwen3.8 27B to 5.9GB, retaining 98.2% of its performance in self-tests.

Beating AI News: AI lab PrismML, co-founded by Caltech mathematician Babak Hassibi, has launched Ternary Bonsai 2 27B, compressing the 27-billion-parameter Qwen3.8 27B model to just 5.9GB. The model quantizes its weights into three values (-1, 0, +1), making its size roughly one-ninth that of the original FP16 version. It supports a 262k-token context window, text-and-image input, and agent tool calls, with its weights released under the Apache 2.0 license. Compared to the first-generation Bonsai 27B launched two months ago, the new version retains nearly the same size but swaps the base model from Qwen3.6 to Qwen3.8. PrismML’s internal tests show its overall benchmark performance retention rate versus the full-size model has risen from approximately 95% to 98.2%, with math and coding results now nearly matching the original. The team also demonstrated running the Cline programming agent and Computer Use directly on an RTX 5090.