PrismML hopes its tiny LLM will change how we all use AI

PrismML released Bonsai 2 27B, a 5.9 GB model that compresses Alibaba’s Qwen 3.8 27B to 9-10× smaller size while matching 98% of its benchmark scores. The startup uses ternary weights (+1, −1, 0) to reduce each weight from 16 bits to three values, enabling high-performance reasoning models to run on PCs and smartphones. Future releases aim to compress several-hundred-billion-parameter models with minimal performance loss.