Frontier AI is getting dramatically cheaper. DeepSeek V4 Flash combines a 284B-parameter Mixture-of-Experts (MoE) architecture, compressed attention, and optimized post-training to deliver frontier-level performance at a fraction of the inference cost. #inference #deepseek #claude #chatgpt #ainews