Comment by ch_sm

you can try ollama, omlx or llama.cpp for instance to download a model and get an inference server running locally. They expose „open ai compatible“ endpoints, so you can configure almost any harness to use them.