DeepSeek V4 drops with 1M context, MIT license

DeepSeek V4 is a masterclass in architectural efficiency that makes frontier-level intelligence affordable by outperforming competitors on cost-per-token rather than just raw scale. The 1.6T parameter MoE architecture of V4-Pro matches or beats Claude 4.6 Max in coding and agentic tasks. DSA and Manifold-Constrained Hyper-Connections solve the long-context cost bottleneck, enabling deep reasoning across massive repositories without exponential cost increases. At 97% lower cost than proprietary competitors, it effectively commoditizes high-end reasoning for production-scale agentic workflows. Hardware optimizations for Huawei's latest chips demonstrate a successful path for AI performance independent of Nvidia's dominance.