Image: Giznova PrismML's Bonsai 2 Squeezes a 27B Reasoning Model Down to 5.9 GB
The Caltech spinout's new release compresses Alibaba's Qwen3.8 27B into a 5.9 GB, phone-friendly footprint while matching 98% of the original's benchmark scores, using ternary weights.
PrismML released Bonsai 2 27B on Thursday, the latest in its family of tiny reasoning models — and it is a striking demonstration of how far model compression has come. Bonsai 2 takes Qwen3.8 27B, a widely used open-source model from Alibaba, and squeezes it down to 5.9 GB. That is a 9x to 10x reduction in memory, small enough to fit on a PC and, PrismML says, possibly a high-end smartphone.
The headline number to scrutinize is the quality retention: PrismML says Bonsai 2 matches 98% of Qwen's aggregate benchmark scores, up from the first Bonsai, released in March, which matched 95% and has since been downloaded more than 11 million times. Benchmark parity is a vendor claim, and as TechCrunch noted, perfect parity is fairly academic — uncompressed models are not that accurate, and benchmarks are not that reflective of real tasks, for a 2% degradation to meaningfully affect real use. The smaller models have racked up another 2.6 million downloads, per the company.
Ternary weights: +1, −1, 0
PrismML says it achieves the compression by shrinking the weights that make up the model — essentially, the information a model learns and stores during training. Normally each weight requires 16 bits. PrismML's approach, which it calls ternary weights, simplifies that down to three values: +1, −1, or 0. With far smaller values to store per weight, the model takes up dramatically less space.
The people behind it are part of the story. PrismML was founded by a group of Caltech researchers and is led by CEO Babak Hassibi, a Caltech professor and an expert in compression technologies. Ion Stoica — Databricks co-founder and director of Berkeley's famed Sky Computing Lab, which birthed technologies from Letta to SGLang — is an adviser. The startup has raised a $22.25 million seed round, backed by Khosla Ventures, Cerberus Capital, and Caltech, and it is reportedly in talks with Apple, though Hassibi declined to comment on that to TechCrunch — treat the Apple rumor as exactly that until confirmed.
Why developers should care
On-device reasoning changes the economics of agentic coding. A capable 27B-class model running on hardware you already own means private inference with no data leaving the device, no API bills, and no latency from round-trips — the exact properties that make always-on coding agents affordable and acceptable inside enterprises. Stoica put it plainly: "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."
Hassibi's next claim is the one to watch: the company plans to apply the compression technique to models in the several-hundred-billion-parameter range within the next couple of months, and he expects it will be easier to retain intelligence at larger sizes — "as a general trend, for larger models, it's easier to get to 100%." If that holds, the frontier may not stay in the cloud. PrismML is not the only lab in this race — Multiverse Computing, founded by a well-known professor from Spain's Donostia International Physics Center, is working on similar compression tech with considerably more funding — but Bonsai 2 is the release that makes the strongest case yet that capable AI will fit in your pocket, not just your data center.