For a user without much technical experience using a ready-made gui like Jan.ai with automatic model download and ability to run models with the ggml library on consumer grade hardware like mac M-series chips or cheap GPUs by either Nvidia or AMD is probably a good start.

For a little bit more technically proficient users Ollama is probably a great choice to start to host your own OpenAI-like API for local models. I mostly run gemma2 or small llama 3.1 like models with that.

Ollama software and you can pick a model that dose what you want. Mistral and Llama are currently best IMO but it changes often.

Oobabooga is a pretty beginner-friendly solution for running LLMs locally. Models are freely available on Huggingface, but look for GGUF quantizations that will fit in your VRAM. The good thing about GGUFs is that they're typically offered in a wide range of sizes so you can pick one that will fit on your GPU. If you use all your VRAM and start offloading to system memory then the generation will be far slower.

I've had the best results with Noromaid20B and Rose20B quants running on a 16GB 4080. Don't expect it to be as smart as GPT 4.0, but those models do a pretty good job of following instruction and writing decent prose.

Once you mess around with Oobabooga a bit, I'd highly recommend picking up the SillyTavern front-end. Oobabooga runs the actual model while SillyTavern manages characters, world lore, and offers a wide range of other features including a "visual novel" mode where you can set up character sprites that emote based on the content of the messages. It takes a while to get the hang of but it's pretty cool.

Nothing special, I got it to give me a recipe for meth and a list of sites to pirate games. Stuff like that

Woo, thanks!