PrismML compresses reasoning model to a size suited for personal devices
AI startup PrismML has released Bonsai 2 27B, a compressed version of Alibaba’s open-source Qwen3.8 27B. The model occupies 5.9 GB, or roughly nine to ten times less memory than the original, which the company says could allow it to run on PCs and potentially high-end smartphones. CEO Babak Hassibi declined to comment on reports that PrismML is in discussions with Apple.
PrismML was founded by Caltech researchers and is led by Hassibi, a professor specializing in compression technologies. Ion Stoica, co-founder of Databricks and director of Berkeley’s Sky Computing Lab, is an adviser. Khosla Ventures, Cerberus Capital and Caltech are among the company’s backers; its seed financing totals $22.25 million.
The company says Bonsai 2 achieves 98% of Qwen’s combined benchmark scores, compared with 95% for the first Bonsai model released in March. That model has been downloaded more than 11 million times, while smaller PrismML models have recorded another 2.6 million downloads. The approach replaces conventional 16-bit model weights with ternary values of +1, −1 or 0. PrismML plans to apply it to models with several hundred billion parameters, while Hassibi acknowledges that compression can still reduce performance.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.