Llama.cpp + Wizard Vicuna (Uncensored, if you want to get the real thing) + one of the web interfaces that are compatible. Should be listed in the readme.
Or try gpt4all which is much easier to use and even offers a selection of downloadable models.
7B/13B/30B+ depends on your hardware, especially GPU.
I use koboldcpp with the vicuna model. Reasonably fast generation (<1 minute) on a 4th gen i7, would probably be on par with chatgpt in terms of speed if you used a GPU.
I believe gpt4all has a self-hostable web interface but I could be wrong. Still it can run on relatively low end hardware (relatively because it still needs a decent amount) and you could just use it on your local computer.