PrismML releases Bonsai 27B, compressing Qwen3.6-27B to 1-bit and ternary weights

PrismML has released Bonsai 27B, a low-bit quantization of the Qwen3.6-27B model that enables running large multimodal models on laptops and phones without retraining.

• Ternary Bonsai 27B uses ternary weights at 1.71 bits per weight for a 5.9GB footprint, retaining 94.6% of the FP16 baseline accuracy.
• 1-bit Bonsai 27B uses binary weights at 1.125 bits per weight for a 3.9GB footprint, retaining 89.5% of the FP16 baseline accuracy.
• The architecture remains unchanged from Qwen3.6-27B, with only normalization and scale parameters kept in higher precision.
• Both variants support a 262K token context window and are available under the Apache 2.0 license for llama.cpp and MLX.

The release addresses memory constraints on edge devices, allowing single-GPU serving on 24GB cards and local inference on mobile hardware while maintaining practical performance.