Well.

Last time round we got our hands on 7B-70b models.

Presumably, a company with as much compute as Meta can train even larger models that weren’t released, that would greatly benefit from the massive global efforts expended for free on the smaller open source LLM family tree of models. And their architecture ends up largely adopted by the open source community, helping build advanced tools to utilize those models.

It seems to me Meta is simply following the old startup tenet: build something people want.

They were well positioned to pull this off with their AI/ML research resources and tradition of openness, and happen to have one of the biggest troves of data.

Maybe there isn’t a benefit to closed-source Llama either.

Community goodwill to offset part of the negativity they deservedly receive for the shady aspects of their main products.

Yes, but also I think it benefits them in the VR market with Quest headsets. It's a small market, but one that they have a huge lead on.