I don’t see why they just don’t allow a smaller model to answer the question while letting the bigger one vet it. The vetting can be asynchronous and can be delivered after a few seconds (if it’s an easy query). If it’s a hard query, the UI can show the answer is currently being vetted or something.