PrismML releases Bonsai 2 27B, a smaller AI model for devices

In short: PrismML says it compressed a popular AI model to 5.9 GB so it can run on a PC and possibly a high-end smartphone.

What happened

PrismML, an AI lab founded by Caltech researchers, released a new model called Bonsai 2 27B. A model is the “brain” behind chatbots and other AI tools.

According to PrismML, Bonsai 2 is a smaller version of Qwen3.8 27B, an open-source model from Alibaba. PrismML says it reduced the model’s memory needs by about 9 to 10 times, bringing it down to 5.9 GB. That size can fit on many computers, and the company says it might also fit on a high-end smartphone.

PrismML says the model keeps most of its quality even after shrinking. It says Bonsai 2 matches 98% of the original model’s combined benchmark scores. Benchmarks are standardized tests for AI (like practice exams), but they do not always match real life use.

The company says it shrinks models by simplifying the “weights,” which are the stored numbers the model uses to make predictions (like the settings on a huge mixing board). PrismML says it turns many weights into just three possible values: +1, -1, or 0. This cuts down how much storage the model needs.

PrismML has raised a $22.25 million seed round, according to TechCrunch, and is backed by Khosla Ventures, Cerberus Capital, and Caltech. TechCrunch also noted a rumor that PrismML is in talks with Apple, which PrismML’s CEO Babak Hassibi declined to comment on.

Why it matters

If advanced AI can run directly on your phone or laptop, it may work faster and keep more of your data on your own device instead of sending it to faraway servers (the cloud). That can be helpful for privacy and for people with limited internet access.

Source: TechCrunch AI