So, I'm relatively new to the scene and I kind of want to do a sanity check.
I've been using gemma-4-26B. Been loving it except for the tool calling unreliability. |

I encountered this fine tune called SuperGemma, that claims to fix these issues.

https://huggingface.co/Jiunsong/supergemma4-26b-abliterated-multimodal

.. But when I run it, it's just.. broken. Often it gives blank responses without an end token. Often it seems to just say things completely unrelated to what I say. The responses that it does seem to give correctly are extremely terse, even when instructed otherwise.

Generally, it's a mess.

So my question is: What the hell? Am I doing this wrong, somehow? This happens even when I run it with the exact settings recommended in the model card. And, obviously, I've had a lot of success with other gemma-4 models.

So is there something I'm missing, here? Please, do tell me if this model works fine for you.