The Mistral stuff is completely proprietary except for the final weights as far as I can tell

Same as most open weight models. There are very few truly open source models, particularly if your definition requires distributing the whole training dataset.

I’d rather resist that since I want to be able to build and modify it myself.

I hope you've got some damn beefy hardware then. Even small models are a pain to train from scratch. You CAN still fine-tune it or train with LoRA.