Key question: is this an entirely new LLM trained from scratch, or is it fine-tuned on top of an existing foundation model like Llama 2 or Mistral?

I think it's a new foundation model, but the announcement doesn't make that clear to me.