PrismML hopes its tiny LLM will change how we all use AI
techcrunch ·
PrismML is betting that capable reasoning large language models do not need to be large. Founded by Caltech researchers and led by compression expert Babak Hassibi, the startup has released Bonsai 2 27B, which compresses Alibaba’s Qwen3.8 27B model to 5.9 GB, reducing its memory requirements by roughly nine to ten times. The model is small enough to run on a PC and potentially on a high-end smartphone. PrismML says its ternary-weight technique reduces model weights to the values +1, 0, or −1, while Bonsai 2 retains 98% of Qwen’s aggregate benchmark performance. The company plans to apply the method to models with several hundred billion parameters, and adviser Ion Stoica says on-device AI could be more accessible and private because it would not need to send data to the cloud. PrismML has raised a $22.25 million seed round and faces competition from other companies developing LLM compression technology.
techcrunch