DeepSeek-V4 Preview Version Officially Launched and Open-Sourced

The DeepSeek-V4, a new model series from DeepSeek, is officially launched today in preview form and simultaneously open-sourced. The series comprises two variants: - **DeepSeek-V4-Pro**, whose performance rivals top-tier closed-source models and leads the open-source domain in agent capabilities, world knowledge, and mathematical/coding reasoning; - **DeepSeek-V4-Flash**, which features a smaller parameter count and delivers faster, more cost-efficient API services. Both models support a 1M (million-token) ultra-long context window and operate in both non-thinking and thinking modes. Technically, DeepSeek-V4 introduces a novel attention mechanism—combining DSA (Dynamic Sparse Attention)—to achieve globally leading long-context capability while significantly reducing computational and GPU memory requirements. For API usage, the `model` parameter accepts `deepseek-v4-pro` and `deepseek-v4-flash`. The legacy model names `deepseek-chat` and `deepseek-reasoner` will be deprecated on July 24, 2026.