Please consider LLaMA.cpp (https://github.com/ggerganov/llama.cpp), which supports a lot of models and doesn't need an expensive GPU.