If the dataset it was trained on is CC-BY-NC, I'm pretty sure the model also has to be CC-BY-NC. However I think this is not respected, or even considered by most people.

I'd go with limiting how competitors can use it as the main deciding factor.