DeepSeek V4.1 Flash claims performance edge over Kimi K3, cuts inference costs

DeepSeek released V4.1 Flash, claiming it outperforms its previous flagship and rivals Kimi K3 on cyber and coding benchmarks, while reducing inference costs and increasing speed. The model uses a 552B-parameter MoE architecture with a new Causal-Encoder-Decoder design. This intensifies China's AI price-performance war, potentially pressuring global AI margins and accelerating commoditization.