Has anyone had success using local LLMs (on like a 12GB GPU) with something like OpenClaw? At least for some subagents. Which models are good?