Ternary Bonsai 2 (27B) just released on Hugging Face. At <6GB in size, it can even run locally in-browser on WebGPU.

Ternary Bonsai 2 (27B), a new language model derived from Qwen3.8-27B, has been released on Hugging Face. This model utilizes ternary weights to significantly reduce its size to less than 6GB, enabling it to run locally in-browser using WebGPU. The architecture of the causal language model remains unchanged from its Qwen3.8-27B origin. A demo is available, showcasing its in-browser capabilities.

Unlike previous models of similar scale, Ternary Bonsai 2 (27B) is the first to achieve local in-browser execution on WebGPU due to its sub-6GB size.