>I heard MoE reduces inference costs Computational costs, yes. You still take the same amount of time for processing the prompt, but each token created through

>I heard MoE reduces inference costs Computational costs, yes. You still take the same amount of time for processing the prompt, but each token created through inference costs less computationally than if you were running it through _all_ layers.