Local LLM runtime. We use it for prompt generation (llama3.1:8b feeds the ComfyUI CLIP prompts) and other tooling that should stay off third-party APIs.