Prism ML introduced the Bonsai 2 27B model with 1.76-bit weights

Company Prism ML released Bonsai 2 27B — a specialized version of the language model using ternary quantization technology. The novelty achieves weight compression up to 1.76-bit, which allows the model to occupy only 5.9 GB of disk space.

Developers claim that Bonsai 2 27B retains 98.2% of the performance of the original Qwen3.8 27B model. Such a high degree of compression provides a 9-fold reduction in memory volume compared to the original architecture, while maintaining functionality for solving programming tasks, computer vision, and autonomous agent operation.

The model's technical specifications are focused on local deployment with high bandwidth. The use of ternary weights reduces hardware requirements, allowing the system to run on consumer equipment without significant loss of answer accuracy. This makes the model suitable for scenarios where energy efficiency and real-time data processing speed are critical.

A discussion of the technical aspects of the implementation and benchmarks is available on the Hacker News platform. This release continues the trend of optimizing heavy LLMs for operation on local devices, minimizing dependence on cloud computing while maintaining a high level of cognitive abilities of the model.