National Supercomputing Internet launches DeepSeek V4 Pro with enhanced agent capabilities

On August 14, the National Supercomputing Internet officially launched the stable version of DeepSeek V4 Pro (DeepSeek-V4-Pro-0813) and the agent framework DeepSeek Harness. The platform, as the first 100,000-card-scale super-AI integrated computing resource pool in China, relies on a self-controlled domestic computing base to provide full-lifecycle support for large model training, fine-tuning, evaluation, and deployment. Users can complete model downloads, private deployment, and agent development in one stop.

DeepSeek-V4-Pro-0813 shows major improvements in agent capabilities, with outstanding performance in production environments, rivaling closed-source models like Claude Fable 5 and Opus-4.8. On August 13, the developer preview (v0.1) of DeepSeek Harness was opened for global testing, with source code released under the MIT license. The supercomputing internet AI community has published the full source code, allowing developers to pull the project files for further development.