It's largely worse from the outside because meta has no goodwill to make mistakes and iterate in public. GPT-2/3 were just as bad at hallucination, if not worse. That doesn't mean that meta isn't iterating on their LLM in private using the 3b users and armies of labelers to fine-tune and improve their model.

From what I understand, they have one of the strongest translation models at scale, and their recommendation models for advertising are likely only rivaled by Google or Bytedance

All those models are only as good as the data they’re trained on. Google and Meta both have the most relevant proprietary data that cannot be rivaled. Not even close.