PrismML's Bonsai 2 27B squeezes a 27B open-source model into 5.9GB

PrismML, an AI lab founded by a group of Caltech researchers, released Bonsai 2 27B on Thursday. The model compresses Qwen3.8 27B, a widely used open-source model from Alibaba, down to 5.9GB.

That is roughly a 9x to 10x reduction in memory versus the original, small enough to fit on a PC and possibly a high-end smartphone. The bet behind the release is that capable, high-performing reasoning models do not, in fact, have to be large.

Accuracy is what PrismML is selling. According to the company, Bonsai 2 matches 98% of Qwen's aggregate benchmark scores, up from 95% for the first Bonsai, which shipped in March. The original model has been downloaded more than 11 million times, and PrismML's even smaller models another 2.6 million, the company says — the trend from one release to the next is the signal it wants people to see.

The technique is about how a model stores what it learned. In a normal model each weight needs 16 bits; PrismML uses "ternary" weights that simplify each one down to +1, -1, or 0. Fewer bits per weight means a smaller footprint.

The company itself is early. It is led by CEO Babak Hassibi, a Caltech professor and an expert in compression technologies, and counts Ion Stoica as an advisor — Stoica co-founded Databricks and directs Berkeley's Sky Computing Lab, which has spun out projects from Letta to SGLang. PrismML is backed by Khosla Ventures, Cerberus Capital, and Caltech, and has raised just a $22.25 million seed round.

It is far from alone in the compression race. Multiverse Computing, founded by a well-known professor from Spain's Donostia International Physics Center, works on the same problem and has raised substantially more money. Hassibi argues PrismML is unique because its models lose virtually no performance compared with the originals, but he also says compression will always have some impact, and whether it ever reaches 100% benchmark parity remains to be seen.

Whether that parity actually matters is a separate question. LLMs are not so accurate in their uncompressed form, and benchmarks are not so perfectly reflective of real tasks, that a 2% degradation would likely change how a model performs in actual use — the surrounding harness matters a lot for accuracy too. The real question is whether on-device models get close enough in daily work that the default place to run a 27B-class model moves from the cloud to the device.

For now, the thing to watch is whether the gap keeps narrowing in the next Bonsai release, and whether the on-device experience matches the benchmark story once real hardware and real workloads are involved.