“Mixture-of-Experts (MoE) architectures enable more capable models that remain efficient at inference time because only a small part of the model's capacity is activated for any given token.”