DeepSeek V3's MoE architecture makes it far more inference-efficient than dense models of comparable scale, enabling low-cost API services