DeepSeek V4 Pro official version, Harness go live on China's National Supercomputing Internet
A person uses DeepSeek app on a mobile phone on Feb. 17, 2025. (Xinhua/Huang Zongzhi)China's National Supercomputing Internet announced on Friday that it had made the official version of DeepSeek V4 Pro (DeepSeek-V4-Pro-0813) and the DeepSeek Harness agent framework available on its platform, a move that a Chinese expert said could lower technical barriers for domestic developers and accelerate the adoption of AI agents in scientific research and industrial applications.DeepSeek-V4-Pro-0813 places a particular emphasis on enhanced agent capabilities. DeepSeek Harness, launched alongside the model and open-sourced under the MIT License, adopts an "everything is a plugin" architecture, allowing components such as models, tools, skills and sessions to be freely replaced and recombined. It offers four operating modes: Standard, PTC, Minimalist and Creative.Backed by an independently developed and controllable national computing infrastructure powered by domestic technologies, the platform enables users to download DeepSeek models and source code, conduct private deployments and distributed inference, and carry out secondary development of AI agents through a one-stop service, helping bridge the critical gap between cutting-edge large-model research and industrial application.According to the platform, the National Supercomputing Internet operates China's first 100,000-accelerator resource pool integrating supercomputing and AI computing, supporting the entire large-model life cycle from training to deployment.DeepSeek-V4-Pro-0813 features enhanced agent capabilities, with particularly significant performance improvements in production environments. Its overall performance is comparable to that of foreign closed-source models such as Claude Fable 5 and Opus 4.8.On August 13, the developer preview of DeepSeek Harness (v0.1) was released for testing by agent-harness developers worldwide, with its source code made available under the MIT License. The National Supercomputing Internet's AI community has also made the complete source code available, allowing developers to download the full project files from its code repository for compilation, debugging, further development and deployment.The National Supercomputing Internet's AI community currently hosts more than 1,700 popular open-source models from China and abroad, including leading Chinese model families such as DeepSeek, Qwen, MiniMax and Kimi.Wang Peng, an associate research fellow at the Beijing Academy of Social Sciences, told the Global Times that the rollout goes beyond adding two new services. It connects an advanced model and agent framework with national computing infrastructure, making computing resources easier to use while strengthening compatibility with domestic systems.The enhanced agent capabilities of DeepSeek V4 Pro reflect a broader shift from models that provide answers to agents that execute tasks. Users can set a goal, while the agent plans the workflow, calls tools and completes the task, expanding AI applications from content generation to software engineering, industrial simulation and scientific computing, Wang said.For research institutions and smaller technology firms, the integrated platform could reduce computing costs and technical complexity. Its 100,000-accelerator resource pool supports the entire model life cycle, while Harness's plug-in architecture simplifies model adaptation, development and deployment.Wang added that the MIT-licensed framework allows flexible customization, enabling smaller teams and individual developers to build AI agents and supporting the growth of China's domestic agent ecosystem.
