“Y Combinator startup Cactus Compute is developing a "hybrid inference" system that uses a local router to direct prompts to either a local model like Jemma or a server-based model like Gemini based on complexity.”