Architecture where a router activates subsets of expert MLPs per token, increasing capacity efficiently.