The problem with LLMs is that they require immense compute power.

To train. But you can run a relatively simple one like phi-3 on quite modest hardware.