Prism ML squeezes a 27B model into a 5.9 GB file
The Pasadena company's ternary flagship keeps more than 98% of its full-precision benchmark scores while shrinking into a ninth of the memory.
Prism ML has squeezed a 27-billion-parameter multimodal model into a 5.9 GB download without, it claims, surrendering much of the quality. The Pasadena company announced Bonsai 2 27B on 17 September, calling it its most capable model yet and a step towards running useful AI agents on the hardware people already own rather than rented data-centre capacity.
What is actually new
Bonsai 2 27B is built on Qwen3.8 27B and ships as a ternary model, which means its weights are stored in three states rather than sixteen bits of precision. The result is a package more than nine times smaller than the full-precision original that still retains over 98% of that model's aggregate benchmark performance, up from the 95% the first Ternary Bonsai 27B managed in July. Prism ML says Bonsai 2 27B is also about 8x faster and uses roughly 5x less energy, and that it handles reasoning, coding, vision, tool calls and long-horizon agentic work.
Why the memory number matters
The workloads worth having in 2026 assume a model can hold a long context, call tools and read images, and that is exactly the blend that normally forces a subscription and a network connection. A 27B-class model in 5.9 GB changes the arithmetic. It fits on a handset or a mid-range laptop, keeps the material being worked on inside the device, and removes the metered per-token cost that makes agents expensive to leave running. The company calls the goal intelligence density: intelligence per bit rather than parameter count.
"Bonsai 27B proved that powerful models do not have to be confined to cloud infrastructure," said founder and chief executive Babak Hassibi. "With Bonsai 2 27B, we are closing the quality gap while keeping the same deployment advantages." Adviser Ion Stoica, a UC Berkeley professor, said the striking part is how little capability is lost given the reduction in footprint, and that if the gap keeps closing it expands where capable models can be deployed.
What to treat carefully
Aggregate retention is the company's own number rather than an independent audit, and averages hide the specific tasks where a compressed model falls apart. Ternary weights also need runtimes that handle them properly, so the speed you get depends on the hardware in front of you, not on a throughput chart. Prism ML is also building on somebody else's open model rather than training its own, so its contribution is compression and packaging, which is where the interesting fight in local AI now sits.
Our opinion
The claim here is not that a compressed 27B model beats a frontier system, because it does not, and anyone selling that line is selling something else. It is that the floor of what runs on a phone has risen to the point where a large slice of everyday AI work no longer needs a data centre at all, and that is a bigger deal for privacy and for bills than another leaderboard win up top. Two caveats are worth carrying. Retention figures measured across a benchmark average say nothing about the long tail, where small models usually wobble first, and a model that fits in 5.9 GB still needs to be loaded, kept warm and fed sensible prompts on a device designed to save battery. Prism ML's real achievement is the compression ratio, and the honest test is not this launch but the next one: if the gap keeps closing, local hardware stops being the compromise.