Not sure if youre asking about already trained models or you want to train yours.
If you just want to have fun the small to medium models are pretty ok. Things like Wizard Vicuna 13b or the smaller 7b. You just have to try some of them until you find ehats best for your use case. Ex I have a model running discord bots (with different personalities) but the same model would work badly with my other projects. Esp considering that with some models you can just chat while others need instructions.
There are also recent models that approach gpt levels. Downside is they are huge in terms of hardware cost (hundreds of gbs of ram, multiple gpus). But they wont necesarly be better than a small more focused model.
Get oobabooga (the automatic1111 of chat llms) and then search for TheBloke on huggingface for models.
So what if you dont agree with their terms? What then? Cant you just host the repo and tell them to fuck off since you sisnt agree to anything?