[numpy-thagoras]

You don't need to load the entire world model in order to be effective at a task. LLM providers already do something similarly described with model routing.