PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B That Run on Laptops and Phones

PrismML has released Bonsai 27B, a low-bit (1-bit and ternary) version of the Qwen3.6-27B large language model, enabling it to run on consumer laptops and smartphones. The release includes two variants under the Apache 2.0 license: a ternary version using {1, 0, -1} weights and a 1-bit binary version. The underlying architecture of Qwen3.6-27B remains unchanged; Bonsai 27B is a compression technique that drastically reduces memory and compute requirements. This allows inference of a 27-billion-parameter model on devices with as little as 8 GB of RAM, previously only feasible on high-end GPUs. The project is open-source, targeting developers and researchers who need on-device AI without cloud dependency. PrismML claims minimal accuracy loss compared to the full-precision model, though independent benchmarks are pending.

Global Impact

Technologically, Bonsai 27B accelerates the trend toward edge AI, reducing reliance on centralized cloud infrastructure and lowering the barrier for AI applications in privacy-sensitive or offline environments. Economically, it could disrupt the cloud AI inference market (worth tens of billions) by enabling local processing, potentially compressing margins for providers like AWS, Azure, and Google Cloud.