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

PrismML simply launched Bonsai 27B. It is a low-bit illustration of Qwen3.6-27B, not a brand new pretrain. The structure is unchanged.

Two variants ship beneath Apache 2.0. Ternary Bonsai 27B makes use of {−1, 0, +1} weights at a real 1.71 bits per weight. Its ideally suited dimension is 5.9GB. 1-bit Bonsai 27B makes use of binary {−1, +1} weights at 1.125 bits per weight, for 3.9GB.

Both are multimodal. The cut up is ~24.8B language weights, a 0.46B imaginative and prescient tower, and 2.5B in embeddings and the LM head. The imaginative and prescient tower is held individually at 4-bit (HQQ). Context is 262K tokens, saved sensible as a result of ~75% of Qwen3.6-27B consideration is linear.