Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

Bonsai 2 27B marks the latest advancement in the Bonsai series of multimodal models, delivering enhanced reasoning, coding, vision, and agentic capabilities while maintaining an exceptionally small computational footprint. Based on the Qwen3.8 27B architecture, this new model implements ternary weights (-1, 0, +1) using FP16 group-wise scaling, achieving 1.76 effective bits per weight and a total footprint of just 5.9GB.

Compared to its full-precision counterpart, Ternary Bonsai 2 27B is nearly nine times smaller, yet retains 98.2% of its aggregate benchmark performance. This represents a near-lossless compression that opens up new possibilities for deploying powerful AI locally. By achieving 98% retention of performance from the full precision Qwen3.8 27B model, Bonsai 2 27B distinguishes itself as a highly capable low-bit alternative.

It excels in coding agents, tool-use systems, multimodal workflows, and long-horizon tasks, which are particularly sensitive to model degradation. The new model's improvements include up to 143 tokens/second on the NVIDIA GeForce RTX 5090 and 46.8 tokens/second on M5 Max, with significantly lower energy consumption of 0.714 mWh/token per token processed.

These gains make Bonsai 2 27B a strong contender for local AI applications demanding both high performance and minimal resource usage. The release of Ternary Bonsai 2 27B signifies a significant step forward in making advanced AI models more accessible and efficient for a wider range of devices and use cases.