The DeepSeek-V4 series represents a significant advancement in large language model architecture, introducing two models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, that natively support a one-million-token context length. To achieve this scale, the researchers developed a hybrid attention mechanism that combines compressed sparse and heavily compressed layers to drastically reduce computational overhead and memory usage compared to previous iterations. Beyond efficiency, the models utilize a novel Manifold-Constrained Hyper-Connections architecture and the Muon optimizer to enhance stability and convergence during the training process. The development pipeline involves specialized domain-expert training followed by a unified distillation process to consolidate capabilities in reasoning, coding, and agentic tasks. Benchmarks indicate that the Pro-Max configuration establishes a new state-of-the-art for open models, rivaling leading proprietary systems in complex reasoning and long-horizon tasks. Ultimately, these innovations provide a foundation for test-time scaling and deeper exploration into intensive, large-scale data analysis