There's a tremendous open source ecosystem in place for Llama already. This is a huge deal for beginners and model integrators alike.
Not to mention that other than beyond 7B models, your options drastically taper off. Mistral and most open base model projects only have models available up to 7 billion parameters or so, which is quite tiny if you are used to the relative ease of using un-finetuned GPT-4 to carry out your tasks of choice.
So what options are there? Falcon 40B and MPT-30B - sure, the weights license is all right, but many in the community have reservations about those models' underperformance, as you can get much more bang for your buck with newer models, in an equal number of weights in a newer base model. Subjectively speaking, it could be a waste of time.
Falcon 180B and Yi 34B weights are both issued under non-free licenses, just like Llama 2.
Is Llama 2 proprietary? For the vast majority of people, for the vast majority of purposes, no. I'm not a lawyer, but I think that Meta would be quite unlikely to do more than cut off your HuggingFace access to the repo where new models will be distributed.
Thanks to Meta for open sourcing Llama!!! Ye sadly the HFF leaderboard doesn't have a high opinion for Falcon. MPT's long context via Alibi did work, just less so when compared to RoPE scaling.
All thanks to Llama - the LLM community is now vibrant and alive!
Llama is generally OSS, except for some gating to large companies - but as a first try I made it work for Llama since the architecture is replicated in other models like Mistral or Yi.
