Eh, the local LLM models (and not just 6B ones but 13B and 30B) out there are actually pretty good these days for chatbots and AI Dungeon-like uses. Maybe not ChatGPT4 quality, but certainly on the level of ChatGPT3 and on normal consumer hardware. Sure, I'm sure there's private cutting edge stuff out there better than this, but most of the benefit isn't that they have better models but that they have large compute farms.