On September 17, Caltech-founded startup PrismML released Bonsai 2 27B, a compressed AI model that fits in 5.9 gigabytes — small enough to run on a laptop or flagship phone. It retains 98% of the performance of Alibaba's full-size Qwen3.8 27B model, which normally requires ten times the memory.
PrismML's trick: "ternary weights." Most AI models store each weight as a 16-bit number. Bonsai simplifies that to three values: +1, −1, or 0. Less data to store, same intelligence. The first Bonsai launched in March and has been downloaded over 11 million times.
Why this matters: you stop being the tenant
Every time you use ChatGPT, your prompt travels to OpenAI's servers. They log it, they learn from it, and you're paying them $20 or $200 a month for the privilege. A 6GB model that runs locally flips the equation. Your data stays on your device. No monthly bill. No one watching your queries. No sudden rule change when the company pivots.
CEO Babak Hassibi says the next models will compress hundreds-of-billions-parameter systems. If they work, the cloud becomes optional. That's the first time in this AI wave that sentence has been true.