Version 1.7.2 adds Gemma 4 12B to the model lineup. It doesn't use the backend's --fit feature, which might be the problem here, if your detected/confirmed VRAM amount is 16+ GB. Hope that fixes it.
