DeepSeek Unveils V4 Preview, Making 1M Long Context Accessible to All
DeepSeek has officially launched and open-sourced the preview version of its latest model series, DeepSeek-V4 . Through structural innovations, this series achieves standardized ultra-long context handling of 1M (one million tokens) and leads domestic and open-source models in Agent collaboration, world knowledge, and logical reasoning.

Two-Version Layout: Pro for Excellence, Flash for Efficiency
DeepSeek-V4 offers two specifications tailored to different application needs:
DeepSeek-V4-Pro (1.6T parameters, 49B activated): Performance rivals top closed-source models. It achieves the best open-source results in Agentic Coding evaluations, with output quality nearing Opus4.6. In math, STEM, and competitive coding tests, it surpasses all publicly evaluated open-source models, demonstrating world-class reasoning capabilities.
DeepSeek-V4-Flash (284B parameters, 13B activated): Prioritizes extreme cost-effectiveness. While its world knowledge is slightly lower than the Pro version, it matches Pro-level reasoning for simple tasks and Agent performance, providing 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 computational and memory demands for ultra-long contexts. This makes 1M context a standard feature across all DeepSeek official services, addressing the industry's high-cost pain point for long text processing.
Deep Adaptation to the Agent Ecosystem
Optimized for mainstream Agent products like Claude Code and CodeBuddy, DeepSeek-V4 supports both non-thinking mode and thinking mode. The API exposes the reasoning_effort parameter, allowing users to adjust thinking intensity (high/max) based on task complexity, significantly boosting performance in complex scenarios such as code generation and document processing.
Access and Open Source Plan
Users can now access the latest model via the official website or the official app, with corresponding API updates. Note that the legacy model names deepseek-chat and deepseek-reasoner will be discontinued after three months (by July 24, 2026).
Open Source Links: Available on Hugging Face and Moba Community .
Technical Report: Published in the Hugging Face repository.
This release of DeepSeek-V4 validates the feasibility of open-source models catching up with top closed-source models in long context and Agent capabilities, laying a solid foundation for AGI popularization through technological breakthroughs.
Related article
Chrome and Edge quietly raise local AI model storage needs to 20GB as browsers evolve into AI inference engines
Google Chrome and Microsoft Edge have raised the disk space requirements for local AI models to 20GB. These browsers silently download the models in the background to enable on-device AI capabilities. In May, Chrome was discovered to automatically do
Zhipu AI Targets $1 Billion ARR With Coding Focus After 5-Month 15x Growth
Multiple independent sources confirm that Zhipu’s Annual Recurring Revenue (ARR) hit $1 billion by July 2026. The company has not yet commented on these reports.AI coding and video generation have emerged as the fastest-growing commercial sectors for
Microsoft Unveils First Cybersecurity AI Model MAI-Cyber-1-Flash, Launches Perception Security Platform
Microsoft has officially introduced MAI-Cyber-1-Flash, its inaugural specialized cybersecurity model, alongside a new AI-driven security platform named Perception. Unveiled at a San Francisco event, these innovations reinforce Microsoft’s standing in
Related Special Topic Recommendations
Comments (0)
0/500
DeepSeek has officially launched and open-sourced the preview version of its latest model series,

Two-Version Layout: Pro for Excellence, Flash for Efficiency
DeepSeek-V4-Pro (1.6T parameters, 49B activated): Performance rivals top closed-source models. It achieves the best open-source results in Agentic Coding evaluations, with output quality nearing Opus4.6. In math, STEM, and competitive coding tests, it surpasses all publicly evaluated open-source models, demonstrating world-class reasoning capabilities.
DeepSeek-V4-Flash (284B parameters, 13B activated): Prioritizes extreme cost-effectiveness. While its world knowledge is slightly lower than the Pro version, it matches Pro-level reasoning for simple tasks and Agent performance, providing faster and more economical API services.
Structural Innovation: DSA Mechanism Enables Long Context Accessibility
Deep Adaptation to the Agent Ecosystem
Optimized for mainstream Agent products like Claude Code and CodeBuddy, reasoning_effort parameter, allowing users to adjust thinking intensity (high/max) based on task complexity, significantly boosting performance in complex scenarios such as code generation and document processing.
Access and Open Source Plan
Users can now access the latest model via the deepseek-chat and deepseek-reasoner will be discontinued after three months (by July 24, 2026).
Open Source Links: Available on
Technical Report: Published in the
This release of
Chrome and Edge quietly raise local AI model storage needs to 20GB as browsers evolve into AI inference engines
Google Chrome and Microsoft Edge have raised the disk space requirements for local AI models to 20GB. These browsers silently download the models in the background to enable on-device AI capabilities. In May, Chrome was discovered to automatically do
Zhipu AI Targets $1 Billion ARR With Coding Focus After 5-Month 15x Growth
Multiple independent sources confirm that Zhipu’s Annual Recurring Revenue (ARR) hit $1 billion by July 2026. The company has not yet commented on these reports.AI coding and video generation have emerged as the fastest-growing commercial sectors for
Microsoft Unveils First Cybersecurity AI Model MAI-Cyber-1-Flash, Launches Perception Security Platform
Microsoft has officially introduced MAI-Cyber-1-Flash, its inaugural specialized cybersecurity model, alongside a new AI-driven security platform named Perception. Unveiled at a San Francisco event, these innovations reinforce Microsoft’s standing in





Home






