I spent the last few days trying to get consistent tool calling out of the new Gemma 4 12b QAT model and had to give up. When the model actually works, it works great, but for my specific use case and workflows it is just not for me. It is a major regression compared to the standard Q5_K_L version, which worked without issue. I know the general consensus is that Qwen is for coding and Gemma is for creatives. But I can tell you for a fact that I code very well with the regular Q5_K_L version. W...