The startup’s latest release, Bonsai 2 27B, compresses Alibaba’s Qwen3.8 27B model into a 5.9 GB package. According to CEO Babak Hassibi, the model captures 98% of the original’s benchmark performance, a jump from the 95% achieved by their initial release in March. While perfect parity remains elusive, the company argues that a 2% variance is negligible for most practical applications. PrismML achieves this efficiency by utilizing 'ternary' weights, which simplify the standard 16-bit data representation down to three values: +1, -1, or 0.
PrismML bets on high-performance AI in your pocket
A modest $22.25 million seed round is fueling a startup aiming to shrink massive reasoning models until they fit comfortably on smartphones and PCs. PrismML, founded by Caltech researchers, claims its latest compression technology retains nearly all the intelligence of its larger predecessors while slashing memory requirements by up to tenfold.
Backed by investors including Khosla Ventures and Cerberus Capital, the firm is already looking toward larger horizons. Hassibi plans to apply this compression to models in the several-hundred-billion-parameter range, where he expects to retain even more intelligence. Advisor Ion Stoica, a co-founder of Databricks, emphasizes that the real value lies in local execution. By moving advanced AI from the cloud to the device, users gain private, cost-free intelligence that runs entirely on hardware they already own.
Comments (0)
No comments yet. Be the first!