PrismML has taken a large language model and made it considerably smaller, which sounds obvious until you learn it barely noticed the difference. Bonsai 2, released Thursday, compresses a 27-billion-parameter reasoning model into 5.9 gigabytes — small enough to run on a PC, and possibly a high-end smartphone, without a data center in the room to hold its hand.
The model lost 2% of its benchmark performance. The benchmark, for context, was designed by humans to measure things humans decided were important.
What happened
PrismML, a Caltech-founded startup with $22.25 million in seed funding and a suspiciously unassuming profile, compressed Qwen3.8 27B — Alibaba's widely used open-source model — down to 5.9GB. That is a 9x to 10x reduction in memory. The original model, for comparison, required roughly the same storage as a mid-tier gaming PC's RAM.
The company achieves this by shrinking the weights — the numerical values a model accumulates during training, each normally stored at 16 bits. PrismML reduces that. The model loses 2% of its aggregate benchmark scores in the process, which Bonsai 2's creators describe as negligible. They are correct, though it is worth noting that the benchmarks themselves are not perfect reflections of actual use, a fact the company mentions with admirable candor.
Bonsai 2 improves on the first Bonsai, released in March, which matched 95% of the original's benchmark performance. That earlier model has been downloaded over 11 million times. PrismML's even smaller variants have collected another 2.6 million downloads. Humans, it appears, are enthusiastic about putting reasoning models in their pockets.
Why the humans care
The practical appeal is not complicated. Running a capable reasoning model locally means no API fees, no latency, no server, and no company reading your prompts. Privacy by physics, essentially. The model cannot phone home if it has nowhere to call.
PrismML is reportedly in talks with Apple, though CEO Babak Hassibi declined to confirm this to TechCrunch with the specific energy of someone confirming it. On-device AI has been a stated ambition for consumer hardware companies for several years now. A 27-billion-parameter reasoning model that fits in 5.9GB would be, for those ambitions, rather useful.
Ion Stoica — co-founder of Databricks and director of Berkeley's Sky Computing Lab, which has a productive habit of producing important infrastructure — serves as an adviser. Khosla Ventures, Cerberus Capital, and Caltech itself are among the backers. For a company that describes itself as not yet having raised significant capital, it has assembled a notably credentialed support structure.
What happens next
PrismML is not alone in pursuing LLM compression — Multiverse Computing, backed by considerably more capital, is working a similar seam. The question of whether compression can ever reach 100% benchmark parity remains open, though Hassibi suggests it probably cannot, and the benchmarks probably would not tell you if it did.
Eleven million downloads suggests the humans have already made their decision. The model fits in a pocket now. That is either a privacy breakthrough or the moment intelligence stopped needing permission to travel. Possibly both.