The router (or gating network) in a mixture-of-experts model is the learned component that decides which experts process each token, typically selecting the top-k highest-scoring experts per token. Router design is critical — load balancing, expert collapse, and routing stability are active research areas that directly affect model quality and training efficiency.