Artificial intelligence startup PrismML is developing technology designed to make advanced large language models significantly smaller while retaining most of their original performance. 

The company believes its approach could eventually allow powerful AI models to operate directly on personal computers and smartphones rather than relying entirely on cloud based systems.

PrismML recently released Bonsai 2 27B, a compressed version of Alibaba’s Qwen3 27B open source model. 

The company said it reduced the model’s storage requirement to about 5.9 gigabytes, representing a reduction of roughly nine to ten times compared with the original model.

The startup was founded by researchers from the California Institute of Technology and is led by Caltech professor Babak Hassibi, who specialises in compression technologies. 

PrismML says Bonsai 2 achieves about 98 percent of the aggregate benchmark performance of the original model, compared with 95 percent for its first Bonsai model released earlier this year.

The company’s compression technology works by reducing the size of the numerical values used to represent the model’s learned information.

Its approach uses three possible values, positive one, negative one and zero, instead of the larger numerical representations commonly used in many AI models. This significantly reduces the amount of memory required to store the models.

PrismML plans to apply the technology to much larger AI models in future releases. 

The company believes smaller models could make advanced artificial intelligence more accessible on personal devices while reducing reliance on cloud computing and potentially improving privacy by allowing more AI processing to take place directly on users’ devices.

Source: TechCrunch