Double the training data, double the trained context (4096 now), a chat tuned varient, the omission of the 35b model for now (it apparently isn’t “safe” enough), and commercial use is allowed (not that most of the people using llama cares about licensing).

There’s a few things, one of the biggest is the commercial license and the 4k context

It also appears on the HF leaderboard now so you can get a very general idea of how (at least the 70b model) compares: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard