Is it free software?

Depends on what you mean. It's fairly straight-forward (and much cheaper) to take an existing model and then fine-tune it. That's how you make domain specific LLMs. It would seem perfectly feasible to make an LLM that was designed to answer questions about a single software project by tuning it with the source for that single project. Whether that would actually give you better results than a generic LLM I don't know. Such an LLM could probably reproduce the original project with higher probability than a generic LLM.

Ideally you'd want a model that understood language, but didn't have any facts other than the ones you specifically give it. But I don't think you can understand a language without knowing something about the real world. I think only the language of math is sufficiently abstract that it could be understood without knowledge of the real world.

IMO the question of whether LLM output is copyrighted will never be a blanket yes/no but depend on the situation. We need to keep an eye on the goals of copyright law (e.g. allowing creators to live of their creations) rather than the machinery (copyright law itself).