PrismML Shrinks 27B AI Model to 5.9 GB With Novel 1-Bit Quantization Technique

Caltech spinoff PrismML unveiled Bonsai 2 27B, compressing a 27-billion-parameter large language model to 5.9 gigabytes using a radical quantization approach that retains 98 percent of the original model's performance. The breakthrough represents a significant milestone in the decade-old quest to run capable AI on consumer devices without relying on cloud compute.