Ternary Bonsai 2
A ternary-weight quantized language model derived from Qwen3.8-27B, compressing a 27-billion-parameter model to under 6 GB. It supports in-browser local inference via WebGPU and uses a {-1, 0, +1} ternary weight representation.
Timeline (last 90 days)
Ternary Bonsai 2 (27B) has been released on Hugging Face.
The model file for Ternary Bonsai 2 is less than 6GB.
Ternary Bonsai 2 can run locally directly in a browser using WebGPU.
Ternary Bonsai 2 is derived from Qwen3.8-27B, a causal language model with a hybrid attention mechanism.
The architecture of Ternary Bonsai 2 itself has not changed; what has changed is the way the weights are represented.
According to the model card, Ternary Bonsai 2 is reduced to one-ninth the size of FP16 while retaining 98.2% of its intelligence performance.
The 98.2% intelligence retention figure comes from the publisher's own model card and has not yet been independently reproduced or verified by a third party.
The publisher has provided a WebGPU-based online demo (webml-community/ternary-bonsai-2-webgpu-kernels).
All Facts (8)
The 98.2% intelligence retention figure comes from the publisher's own model card and has not yet been independently reproduced or verified by a third party.
50%UnverifiedThe model file for Ternary Bonsai 2 is less than 6GB.
50%UnverifiedTernary Bonsai 2 can run locally directly in a browser using WebGPU.
50%UnverifiedTernary Bonsai 2 is derived from Qwen3.8-27B, a causal language model with a hybrid attention mechanism.
50%UnverifiedThe architecture of Ternary Bonsai 2 itself has not changed; what has changed is the way the weights are represented.
50%UnverifiedAccording to the model card, Ternary Bonsai 2 is reduced to one-ninth the size of FP16 while retaining 98.2% of its intelligence performance.
50%UnverifiedThe publisher has provided a WebGPU-based online demo (webml-community/ternary-bonsai-2-webgpu-kernels).
50%UnverifiedTernary Bonsai 2 (27B) has been released on Hugging Face.
50%