National Supercomputing Internet Launches DeepSeek-V4-Pro-0813 and DeepSeekHarness Agent Framework

On August 14, the National Supercomputing Internet announced the official launch of DeepSeek-V4-Pro-0813 and its intelligent agent framework, DeepSeekHarness. This platform represents China’s first large-scale integrated computing resource pool with a capacity of over 100,000 AI accelerator cards, offering full-lifecycle computing support for large models to research institutes, innovative enterprises, and technical developers. Built on domestically developed computing infrastructure, the platform enables users to download DeepSeek-series models and source code, deploy them privately, perform distributed inference, and conduct secondary development of intelligent agents. DeepSeek-V4-Pro-0813 features enhanced agent capabilities and delivers significantly improved performance in production environments, matching the overall performance of leading closed-source international models such as Claude Fable5 and Opus-4.8. On August 13, the DeepSeekHarness developer preview (v0.1) was open-sourced under the MIT license and made available globally for testing. The full source code of DeepSeekHarness is now accessible on the Supercomputing Internet AI Community, allowing developers to pull the entire project repository with a single click for compilation, debugging, iterative development, and deployment.