Model Releases
DeepSeek-V4-Pro: The New Open-Source King of AI Agents
DeepSeek-V4-Pro has officially arrived, bringing a massive 1.6T MoE architecture and unprecedented agentic capabilities to the open-source ecosystem.
13 августа 2026 г.
Model ReleaseDeepSeek-V4-Pro
Introduction: A Milestone in Open-Source AI
On August 13, 2026, the AI landscape shifted once again. DeepSeek has officially released DeepSeek-V4-Pro (version V4-Pro-0813), marking the transition from a highly anticipated preview to general availability. This isn't just another incremental update; it is a milestone release that positions DeepSeek as a direct competitor to the most advanced closed-source models in the world.
For developers and AI engineers, the significance of V4-Pro lies in its dual nature: it offers the raw power of a massive 1.6T parameter model while maintaining the accessibility of an open-weights release. Whether you are accessing it via the official web interface, mobile app, or high-performance API, the V4-Pro architecture is designed to handle the most complex reasoning and autonomous agent tasks ever seen in an open-source model.
- Official Release Date: August 13, 2026
- Model Version: V4-Pro-0813
- Status: General Availability (GA)
- Open Source: Weights available on Hugging Face
Architecture: The Power of 1.6T MoE
At the heart of DeepSeek-V4-Pro is a sophisticated Mixture-of-Experts (MoE) architecture. The model boasts a staggering 1.6 trillion total parameters, but thanks to efficient MoE routing, only 49 billion parameters are active during any single inference pass. This balance allows for 'frontier-class' intelligence without the prohibitive latency typically associated with trillion-parameter dense models.
A standout feature for engineers is the introduction of 'Thinking-effort' levels. Users can now explicitly control the computational budget for a response by selecting between low, high, and max effort levels. This is available for both V4-Pro and the lighter V4-Flash, allowing developers to optimize for either speed or deep reasoning depending on the specific use case.
- Total Parameters: 1.6T
- Active Parameters: 49B (MoE)
- Thinking Levels: Low, High, Max
- Native Responses API: Supported with one-click Codex configuration
Agentic Performance: Breaking Benchmarks
DeepSeek-V4-Pro is explicitly engineered for the 'Agent Era.' While previous models excelled at chat, V4-Pro is built to use tools, navigate repositories, and execute complex workflows. The benchmark results are nothing short of historic, particularly in areas where traditional LLMs struggle, such as long-term planning and tool integration.
In the HLE (Humanity's Last Exam) benchmark, the model achieves a score of 42.7, which jumps to an incredible 60.0 when equipped with tools. Furthermore, its performance in specialized environments like Terminal Bench 2.1 (87.9) and Cybergym (83.3) demonstrates a level of technical proficiency that rivals the best proprietary coding assistants currently on the market.