PrismML releases Bonsai 27B, a 1-bit dense LLM running in the browser

The PrismML team has released Bonsai 27B, a large language model designed to run locally in a web browser using custom WebGPU kernels. The model utilizes 1-bit quantization to significantly reduce its memory footprint while maintaining high performance.

• The 1-bit quantization shrinks the model size from 54GB to just 3.8GB, a reduction of 93%.
• Despite the aggressive compression, the model retains 90% of its original intelligence.
• A collection of checkpoints is available on Hugging Face under the PrismML organization.
• An interactive demo utilizing WebGPU kernels is hosted on Hugging Face Spaces.

This release enables users to run a large-scale dense model directly in their browser without requiring high-end local hardware.