ProCreations released an updated model Ternary-Bonsai-2-27B-MTP

Developers presented Ternary-Bonsai-2-27B-MTP — an updated version of the language model that has undergone fine-tuning. The release is focused on improving computational performance and optimizing the architecture for specific hardware environments.

The key change was the introduction of Q8_0 head, which made it possible to achieve a speed increase of 1.26%. The model requires the use of a specialized patched native runtime for Linux x86-64 systems. For correct system operation, a CUDA 13.3 environment is required.

During testing, the model demonstrated a result of 16 points on the internal metric in 3 hours of data processing. The use of Ternary-Bonsai-2-27B-MTP assumes the presence of a tuned environment that meets the requirements for NVIDIA libraries and processor architecture.

This release continues the series of optimizations for the Bonsai model family, aimed at reducing inference latency. The update allows developers to integrate the model into high-load systems where compatibility with the latest driver and runtime versions is critically important.