A 27B Model That Fits on a Phone
PrismML, a startup founded by Caltech researchers and led by professor Babak Hassibi, released Bonsai 2 27B on Thursday — a compressed version of Alibaba's widely-used Qwen3.8 27B that shrinks the model from roughly 54 GB to just 5.9 GB. The result: a reasoning-capable LLM small enough to run locally on a PC or potentially a high-end smartphone, retaining 98% of the original's aggregate benchmark scores.
The key innovation is ternary weight compression. Where traditional models store each weight in 16 bits, PrismML reduces each to one of three values: +1, -1, or 0. This 9x-10x memory reduction trades minimal benchmark fidelity for massive portability gains — a 2% performance drop that is unlikely to matter in practical use, given that even uncompressed LLMs are imperfect.
Traction Already There
The original Bonsai, released in March, matched 95% of benchmarks and has been downloaded over 11 million times. PrismML's smaller models have been downloaded another 2.6 million times. The startup raised a $22.25 million seed round backed by Khosla Ventures, Cerberus Capital, and Caltech.
The Bigger Play: Hundreds of Billions of Parameters
Hassibi told TechCrunch the next releases, expected within months, will target models in the several-hundred-billion-parameter range, where he expects compression to be easier to achieve. As model size grows, there is more room to compress without losing intelligence, he said.
Adviser Ion Stoica — co-founder of Databricks and director of Berkeley's Sky Computing Lab — frames the significance in consumer terms: You are going to have intelligence at your fingertips, and it's going to be free because it's going to run on the device you already bought. It's also going to be private, because you're not going to send it to the cloud.
What It Means
If compression quality scales with model size as Hassibi predicts, the economics of AI inference change fundamentally. Cloud GPU costs — currently the bottleneck for AI deployment — become optional rather than mandatory for many use cases. Local processing also sidesteps the privacy concerns that have slowed enterprise and healthcare adoption of cloud-based AI.
PrismML is not alone in this space — Spain's Multiverse Computing has raised significant capital for similar work — but its benchmark results and download traction suggest a technical lead in ternary-weight compression specifically.
