Caltech Advances AI Model Compression, Running 1-Bit Model on Smartphones

A California Institute of Technology team has developed a model-compression technique that represents each weight with just 1 bit, targeting the memory and power constraints of large language models. The advance could lower barriers to on-device deployment, allowing models to perform inference on mobile devices such as iPhones without relying entirely on cloud computing while preserving their core capabilities.

The team recently founded a startup called PrismML and open-sourced Bonsai 8B, an 8-billion-parameter model. The release demonstrates that a 1-bit model can run smoothly on a smartphone while sharply reducing memory requirements and energy consumption. Available information does not specify when the technology was unveiled or PrismML was founded, nor does it disclose any funding or investment amounts.