Your RTX 3090 can run better local models than a painfully compressed 70B. We ranked the best 24GB LLMs for coding, chat, agents, and reasoning, with one clear winner. #AI