AI firm PrismML has recently enhanced its Bonsai 2 model lineup by introducing three cutting-edge ternary weight models, built upon the Qwen3.5 architecture: Ternary Bonsai 2 2B, 4B, and 9B. These new additions, complementing the previously released 27B variant, are specifically designed to meet the diverse demands of edge deployment, ranging from simple to intricate tasks. The models utilize {-1, 0, +1} encoded parameters, resulting in remarkably compact storage requirements of approximately 0.43GB, 0.95GB, and 2GB, respectively. This represents a more than 9-fold reduction in size compared to their full-precision counterparts.

In terms of performance enhancements, these models exhibit notable improvements in reasoning capabilities, visual comprehension, tool utilization, and adherence to instructions. Particularly, the 9B model showcases a significant reduction in token consumption, by about 40%, during standard instructional and everyday math tasks.

When deployed on hardware such as the RTX 4090 and M5 Pro, the new models deliver up to 5 times the throughput of full-precision models, while simultaneously enhancing energy efficiency by 3 times. On mobile devices like the iPhone 18 Pro Max, the 9B model achieves an impressive rate of up to 34 tokens per second, and the 2B model reaches up to 120 tokens per second, ensuring swift and responsive performance.

The model weights are openly licensed under the Apache 2.0 license, facilitating seamless deployment through platforms like llama.cpp on NVIDIA GPUs or MLX on Mac, iPhone, and iPad, thereby broadening their accessibility and usability across various devices and environments.