ProCreations released the Ternary-Bonsai-2-27B-DFlash2 model

Developers of Ternary-Bonsai-2-27B-DFlash2 presented an updated version of the model that has undergone further training using r2-dflash weights. The release is focused on improving performance while maintaining the architectural parameters of the base model.

The model is available in BF16 format for training and Q8_0 for inference. During testing, a speed generation increase of 0.54% was recorded according to the token decoding per second metric. The update includes a patch for native runtime, ensuring compatibility with Linux x86-64.

The technical implementation supports work with CUDA 13.3 and is optimized for the SM120 Blackwell architecture. The model demonstrated a result of 15 points on the internal evaluation system for 3 hours of data processing.

This release expands the Ternary-Bonsai line, offering developers tools for working with the latest hardware. The use of specific dflash weights allows for more precise tuning of the model's behavior in tasks with high computational resource loads.