Understand how this architecture uses a single-expert routing mechanism to scale models to trillions of parameters efficiently.