PrismML Bonsai 2 27B: A 5.9 GB Reasoning Model That Fits on Your PC

PrismML just released Bonsai 2 27B, a compressed reasoning model that shrinks Alibaba's Qwen3.8 27B down to 5.9 GB. That is small enough to run on a PC or a high-end smartphone, with a claimed 98% of the original's aggregate benchmark scores. For builders shipping edge AI, agentic workloads, or privacy-sensitive applications, this is the most practical on-device reasoning model yet.

What Bonsai 2 27B actually is

Bonsai 2 27B is a ternary-weight version of Qwen3.8 27B, an open-source model from Alibaba. PrismML compresses each weight from 16 bits down to one of three values: +1, -1, or 0. This "ternary" approach reduces memory by about 9x to 10x compared to the original, bringing the model from roughly 54 GB down to 5.9 GB. The company says the model retains 98% of Qwen's aggregate benchmark performance, up from 95% in the first Bonsai release in March.

The startup, founded by Caltech researchers and led by compression expert Babak Hassibi, counts Ion Stoica (Databricks co-founder and director of Berkeley's Sky Computing Lab) as an adviser. It has raised a $22.25 million seed round from Khosla Ventures, Cerberus Capital, and Caltech.

Why the compression technique matters for builders

Most on-device models today are tiny 1B-7B parameter models that sacrifice reasoning and coding capability. Bonsai 2 27B keeps a 27B parameter footprint but fits into consumer hardware. That changes the trade-off for builders who want capable reasoning, agentic planning, or code generation without shipping data to the cloud.

Ternary weights are the key. Instead of storing each parameter as a 16-bit or 32-bit floating point number, the model stores only the sign and zero. This is far more aggressive than standard quantization (4-bit or 8-bit), but PrismML's results suggest the accuracy loss is modest for most tasks. The company has already seen over 11 million downloads of the original Bonsai and 2.6 million for smaller variants, indicating real-world usage.

Practical implications: lower latency, no API costs, privacy

Running a 27B reasoning model locally means zero per-token API fees, no network latency, and no data leaving the device. Ion Stoica described it as "intelligence at your fingertips ... free because it's going to run on the device you already bought" and private because data is not sent to the cloud.

For agent builders, this opens up use cases where cloud round-trips are too slow or where users demand data residency. PrismML is rumored to be in talks with Apple about integrating the technology into iPhones, though CEO Hassibi declined to comment. If a 27B reasoning model can run on a phone, that changes what on-device AI assistants and agents can do.

What to watch: caveats and next steps

PrismML does not claim perfect parity. Hassibi acknowledged that compression will always have some impact, and the 98% figure is an aggregate across benchmarks. Individual tasks may vary, and the harness around the model matters for accuracy. The company's next target is several-hundred-billion-parameter models, which Hassibi expects to be "easier" to compress without losing intelligence.

Competitors like Multiverse Computing are also working on LLM compression, so PrismML is not alone. But the combination of strong Caltech research roots, iterative improvements from 95% to 98% parity, and potential Apple interest makes Bonsai 2 27B a model worth testing for any builder evaluating on-device reasoning.