PrismML released compressed model that runs on a laptop and nears original performance

Compression without losing intelligence

AI lab PrismML released Bonsai 2 27B yesterday, a compressed version of Alibaba's Qwen3 27B that compresses the 27-billion-parameter model down to a 5.9 GB footprint — a 9–10× memory reduction that lets the model run on a personal computer and potentially on a high-end smartphone. The company says the compressed version hits 98% of the original's aggregate benchmark scores, up from 95% for the first version released in March.

Ternary-weight technique

The gain comes from a method PrismML calls "ternary weights." Instead of storing each weight in 16 bits, the approach restricts values to three options: +1, −1, or 0. Weights are the information the model learns during training, and collapsing their representation slashes storage dramatically. The full technical write-up is on the project's Hugging Face page.

Caltech roots, Berkeley advisor

PrismML was founded by researchers from the California Institute of Technology and is led by Babak Hassibi, a Caltech professor and compression specialist. Its advisor is Ion Stoica, co-founder of Databricks and director of the Sky Computing Lab at Berkeley, an incubator that has produced Letta and SGLang among others. The seed round totals $22.25 million from investors including Khosla Ventures, Cerberus Capital, and Caltech itself.

Millions of downloads, but the road to 100% is long

The first Bonsai release has logged more than 11 million downloads, with smaller variants adding another 2.6 million. Hassibi acknowledges that compression always exacts some performance cost, and that benchmarks don't fully capture real-world tasks, so a 2% gap probably won't materially change the user experience. He also notes that the inference harness matters for accuracy at least as much as the weights themselves.

Next target: hundreds of billions of parameters

The next step is applying the same technique to models in the hundreds of billions of parameters, with a launch expected in the coming months. Hassibi believes larger models may actually be easier to compress without losing intelligence because "there's more room to compress without losing." Stoica adds that the direction puts intelligence at your fingertips, for free, because it runs on hardware you already own. Rumors of talks with Apple drew no comment.