DeepSeek V3 and R1, leveraging the MoE architecture, compress inference costs to a fraction of those of traditional dense models while maintaining top-tier performance.