This gemma4 model was trained 2x faster with
Unsloth and Huggingface's TRL library.
GemPT-26B-A4B is a fine-tuned/distilled variant of Gemma-4-26B-A4B-IT, customized to enhance the original model's reasoning capabilities by injecting step-by-step CoT extracted from GPT-5.4-High.
It was trained on a single MI300X accelerator, using LoRA.
