It can't run LLMs very well, you'll be limited to tiny models with no coding ability and they'll be slow.