PrismML argues that high-performing, reasoning-capable large language models do not necessarily have to be large. Founded by Caltech researchers, the company aims to make these models small enough to fit on PCs and smartphones. The company’s CEO is Caltech professor Babak Hassibi; Ion Stoica, a co-founder of Databricks and director of the Berkeley Sky Computing Lab, also serves as an adviser. PrismML is backed by Khosla Ventures, Cerberus Capital and Caltech. The company’s seed funding to date stands at $22.25 million.
PrismML released the Bonsai 2 27B model on September 17, 2026. The model compresses Alibaba’s open-source Qwen3.8 27B model to a size of 5.9 GB. This means a 9x to 10x reduction in memory compared with the original model, allowing it to fit on a PC and possibly a high-end smartphone. Hassibi declined to comment on claims that the company may be in talks with Apple.
Bonsai 2 reaches 98% of Qwen’s overall benchmark scores. The figure was 95% for the first Bonsai model, released in March. According to the company, the first model has been downloaded more than 11 million times, while smaller PrismML models have been downloaded 2.6 million times. Hassibi says compression may not preserve performance entirely, but that a 2% loss may not be decisive in real-world use.
PrismML reduces the “weights” that carry the information models learn during training. While each weight normally requires 16 bits, the company’s “ternary” approach reduces them to three values:
- +1
- −1
- 0
The company plans to apply the same method to larger models in the several-hundred-billion-parameter range over the next few months. Hassibi said that preserving intelligence while compressing large models may be easier. According to Stoica, the technology could enable advanced models to run on users’ devices, making them free and more private because no data would be sent to the cloud.
Why it matters
This approach raises the possibility that advanced language models may not require a constant cloud connection and powerful data center infrastructure to run. Running models on users’ devices could create a different mode of use in terms of privacy, since data would not be sent to the cloud; however, it is unclear on which devices this would be possible, and at what speed and energy consumption. The limited loss in benchmark results does not eliminate the need to assess to what extent compressed models will remain sufficient for practical use. Applying the ternary method to larger models will show whether the technology can extend beyond a single model. The central question, therefore, is how memory savings will be balanced with performance and ease of use in real-device experiences.