DeepSeek has announced the official release and open-source of the preview version of its new series model DeepSeek-V4. This series of models achieves standardized ultra-long context capabilities of 1M (one million words) through structural innovation, and has reached leading levels in Agent collaboration, world knowledge, and logical reasoning performance within the domestic and open-source fields.

DeepSeek

Two-Version Layout: Pro for Excellence, Flash for Efficiency

DeepSeek-V4 is divided into two specifications based on application needs:

  • DeepSeek-V4-Pro (1.6T parameters, activated 49B): Performance reaches top closed-source models. In the Agentic Coding evaluation, it achieves the best open-source level, with delivery quality close to Opus4.6; in math, STEM, and competition code evaluations, it surpasses all publicly evaluated open-source models, demonstrating world-class reasoning performance.

  • DeepSeek-V4-Flash (284B parameters, activated 13B): Focuses on extreme cost-effectiveness. Although its world knowledge is slightly less than the Pro version, it matches the Pro version in reasoning ability for simple tasks and Agent performance, offering faster and more economical API services.

Structural Innovation: DSA Mechanism Enables Long Context Accessibility

DeepSeek-V4 introduces a pioneering DSA sparse attention mechanism. By compressing at the token level, the model significantly reduces the computational and memory demands of ultra-long context. This means that 1M context will become a standard feature across all DeepSeek official services, solving the industry's pain point of high costs for long text processing.

Deep Adaptation to the Agent Ecosystem

For current mainstream Agent products (such as Claude Code, CodeBuddy, etc.), DeepSeek-V4 has been specially optimized. The model supports non-thinking mode and thinking mode, and opens up the reasoning_effort parameter in the API, allowing users to adjust the intensity of thinking based on task complexity (high/max), significantly improving performance in complex scenarios such as code generation and document processing.

Access and Open Source Plan

Currently, users can experience the latest model through the official website or the official app, and the API has also been updated accordingly. Notably, the old model names deepseek-chat and deepseek-reasoner will be discontinued after three months (July 24, 2026).

  • Open Source Links: Have been launched on Hugging Face and Moba Community.

  • Technical Report: Has been disclosed in the Hugging Face repository.

This release of DeepSeek-V4 not only verifies the feasibility of open-source models catching up with top closed-source models in long context and Agent capabilities, but also provides a solid foundation for the popularization of AGI through technological breakthroughs.