The only problem I had with Qwen, fine tuning on Colab, it takes 31 t/s while Gemma 4 is around 9 t/s, otherwise, one of best local LLM