Ternary Bonsai 2 27B Model Offers Near-Lossless Compression with 9x Smaller Footprint

Introduction of Ternary Bonsai 2 27B

The Ternary Bonsai 2 27B model has been released, building upon previous Bonsai 27B models. This new iteration aims to provide stronger reasoning, coding, vision, and agentic capabilities within a significantly reduced memory footprint, enabling efficient local deployment and better energy efficiency.

Technical Specifications and Compression

Ternary Bonsai 2 27B utilizes ternary {-1, 0, +1} weights with FP16 group-wise scaling, resulting in 1.76 effective bits per weight. This compression method leads to a total model footprint of 5.9GB. The model supports a 262K-token context window and multimodal text-and-image input, and is released under the Apache 2.0 license.

Performance and Capability Retention

Compared to its full-precision equivalent, Ternary Bonsai 2 27B is over 9 times smaller. Despite this reduction in size, it maintains 98.2% of the aggregate benchmark performance. This level of retention allows for nearly identical capability in a footprint that can run in more diverse local environments.

Improvements Over Previous Version

This new release improves upon the first Bonsai 27B by incorporating a stronger base model, Qwen3.8 27B. It also achieves a higher aggregate capability retention of 98.2% against the full-precision model and shows improved performance in reasoning, coding, vision, and long-horizon agentic tasks.

Benchmark Results

Across a suite of benchmarks covering reasoning, math, coding, instruction following, vision, and agentic tool use, Ternary Bonsai 2 27B scored 83.9. This score indicates it retains 98.2% of Qwen3.8 27B’s aggregate performance. The retention of capability is particularly noted in areas sensitive to model degradation, such as coding agents and multimodal workflows.