PrismML's Bonsai 2 compresses Alibaba Qwen3.8 27B to 5.9GB
PrismML, a startup from Caltech researchers, just dropped Bonsai 2, a new AI model that shrinks massive language models down to a size small enough to fit on a PC and, possibly, a high-end smartphone.
They managed to take Alibaba's huge Qwen3.8 27B model and compress it down to just 5.9GB, and it still works almost as well as before.
Bonsai 2 uses ternary +1/0/-1 weights
Bonsai 2 uses a clever ternary weights trick, where each weight is simplified to +1, 0, or -1 instead of using lots of numbers. This keeps things light without losing much quality.
Backed by big names like Khosla Ventures, Cerberus Capital, and Caltech, and adviser Ion Stoica, a co-founder of Databricks and the director of Berkeley's Sky Computing Lab, PrismML wants to use this method on even bigger models soon, potentially making advanced AI tools way more accessible for everyone.