PrismML shrinks reasoning models to fit on phones and PCs

PrismML, an AI lab founded by a group of Caltech researchers, released Bonsai 2 27B on Thursday, its latest model in a family built to prove that capable, reasoning large language models don’t have to be large. The model compresses Qwen3.8 27B, a widely used open-source model from Alibaba, down to 5.9 GB — small enough to fit on a PC and, possibly, a high-end smartphone, and a 9x to 10x reduction in memory versus the original.

The startup is led by CEO Babak Hassibi, a Caltech professor and expert in compression technologies, and counts Databricks co-founder and Berkeley Sky Computing Lab director Ion Stoica as an adviser. PrismML has raised just a $22.25 million seed round so far and is backed by Khosla Ventures, Cerberus Capital, and Caltech. It is rumored to be in talks with Apple, though Hassibi declined to comment on that to TechCrunch.

Hassibi says Bonsai 2 matches 98% of Qwen’s aggregate benchmark scores, up from 95% for the original Bonsai released in March. That first model has already been downloaded over 11 million times, and PrismML’s smaller models have added another 2.6 million downloads, the company says. PrismML isn’t the only company chasing LLM compression — Multiverse Computing, founded by a professor from Spain’s Donostia International Physics Center, is another, and has raised far more funding.

The compression technique, which PrismML calls “ternary” weights, reduces each weight in a model from the usual 16 bits down to just three possible values — +1, -1, or 0 — cutting the space a model needs to store its parameters. Hassibi says the next models, hopefully in the next couple of months, will be in the several-hundred-billion-parameter range, and that larger models should be easier to compress without losing intelligence.

Why it matters

Stoica says the appeal of the approach is that it puts capable models directly on a device people already own — for free, and without sending data to the cloud.