“The mixture of experts architecture provides efficiency wins for both foundation models like GPT-4 and reasoning models.”