What about Meta AI? For reasons I cannot comprehend they have been releasing quality research for free for years like PyTorch, FastText vectors, and the LLaMa models.
I don't know how to reconcile the one side of the company that should be burnt to the ground and the one that's pushing local models forward, but I'd say it's worth considering.
At FAANG, open source is de rigeur for things you can’t make money off of, either because it’s an ecosystem play or someone asked their boss.
You’d be surprised how little drama there is around this.
I’d note that the department that made open LLMs hasn’t produced any work since they produced a Gemini 2.5 Flash equivalent with much better tool calling, because the God King threw a fit. Without reasoning. And they had a reasoning model on deck that was cancelled too.
I don't know how to reconcile the one side of the company that should be burnt to the ground and the one that's pushing local models forward, but I'd say it's worth considering.