I made a little comparison test, your model is included
Hey.
Very interesting model. In practice, it's a really nice change of pace compared to Gemma 4's normal vocabulary. Really fun to play with.
I'm running a comparison of different models on a task that LLM are not trained on, and I compare models based on said task. They get an image, basic drawing tools and are told to "redraw" the image to the best of their abilities. That's completely out of any training data models are trained on. Out of scope results are a topic I'm very interested in. It doesn't prove that your model is good/bad for its intended purpose, but it gives a good comparison point to the original one.
Anyway, your changes to the model, while I didn't notice anything really missing when talking to it, are very obvious during this test. The thing you touched had a very clear role in function calling and in spatial awareness. I perfectly get it's not the goal of your model (actually the test is really outside of any model's goals), but I find the result kinda informative on its own.
It was done on an heretic version of it, but informal tests on your own directly didn't seems to really change much, which is consistant with other heretic vs official comparisons I made.
Anyway, here's the link to my post.
https://huggingface.co/SerialKicked/ModelTestingBed/discussions/4
And again, thanks for the model, had a lot of fun with that one!
Cheers!