A new startup called PrismML is developing technology that reduces the size of large language models enough to run on personal computers and high-end smartphones. Founded by Caltech researchers and led by professor Babak Hassibi, the company released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B model that occupies just 5.9 gigabytes of memory — a reduction of nine to ten times compared to the original. This allows the model to operate on standard consumer hardware without requiring cloud servers.

PrismML's compression method, known as ternary weighting, simplifies the numerical weights within a model from 16-bit precision to one of three values: positive one, negative one, or zero. This approach drastically reduces storage requirements while maintaining nearly identical performance, according to internal benchmarks. Bonsai 2 matches 98% of Qwen's aggregate benchmark scores, up from 95% for the previous version, Bonsai 1, which has already been downloaded over 11 million times. Smaller variants of the model have also been downloaded an additional 2.6 million times.

The company emphasizes that its technique preserves nearly all of the original model's capabilities, with only minor degradation that is unlikely to affect real-world usability. Hassibi notes that as models grow larger, there is more potential to achieve full benchmark parity through compression, suggesting that future releases may approach 100% performance retention.

Ion Stoica, a Databricks co-founder and advisor to PrismML, highlighted the broader implications of this technology, stating that it could make advanced AI models more accessible and private by allowing them to run locally on user devices. He described the development as a step toward "intelligence at your fingertips" that does not rely on cloud infrastructure.

PrismML has raised $22.25 million in seed funding from investors including Khosla Ventures and Cerberus Capital, and continues to build on research originally conducted at Caltech. The company has not confirmed reports of discussions with Apple, and has directed inquiries to its technical documentation, including a Hugging Face page detailing the compression methodology.