Major Upgrade for National Supercomputing Internet: DeepSeek V4 Pro and Harness Framework Now Live

The National Supercomputing Internet platform has rolled out a significant update, deploying the stable version of DeepSeek V4 Pro large language model alongside a new agent development framework called Harness. This move represents a new phase in integrating national computing resources with advanced AI technologies.

100,000-Card Computing Pool Opens Access

At the core of this upgrade is China's first 100,000-card level super-intelligent computing resource pool. This unprecedented computing cluster will provide support to three main user groups:

  • Research Institutions: Stable and reliable computing power for cutting-edge AI research
  • Technology Innovation Companies: Reduced technical barriers and costs for model training and deployment
  • Developer Communities: Full lifecycle services from model fine-tuning to application deployment

This centralized computing supply model addresses common challenges in AI development, including fragmented resources and low utilization rates.

V4 Pro Focuses on Enhanced Agent Capabilities

The official release of DeepSeek V4 Pro (version V4-Pro-0813) features targeted improvements over previous iterations. Development teams concentrated on enhancing agent capabilities, resulting in better performance in task planning, tool utilization, and multi-turn conversations.

The simultaneously launched Harness framework adopts a fully open-source approach under the MIT license, allowing developers free use, modification, and distribution—significantly lowering adoption barriers.

Modular Architecture Enables High Flexibility

The most notable feature of the Harness framework is its "everything as a plugin" design philosophy. This modular approach allows various agent components to function independently:

  • Base models can be swapped according to task requirements
  • External tools and APIs integrate as plugins
  • Domain-specific skill modules support plug-and-play functionality
  • Conversation management and memory components allow custom configuration

This design enables developers to construct and adjust agent systems like building blocks, quickly assembling needed functions for specific application scenarios.

Four Operation Modes for Diverse Needs

To accommodate varied development requirements, the Harness framework offers four preset operation modes:

Standard Mode suits most general tasks, balancing performance and resource consumption. PTC Mode (Professional Tuning Configuration) targets specialized scenarios requiring precise control. Minimalist Mode prioritizes response speed and lightweight deployment, while Creative Mode provides greater freedom for exploratory research and innovative applications.

This layered design allows the framework to serve both beginners seeking quick onboarding and professional teams with advanced performance and control requirements.

Industry Impact and Future Prospects

This upgrade to the National Supercomputing Internet represents more than just product releases—it signifies the formation of new research infrastructure. By integrating top-tier models, open-source frameworks, and national-level computing resources, the platform is building a comprehensive AI development ecosystem.

For China's AI industry, this integrated "model + framework + computing" solution may accelerate the adoption and innovation of agent technologies. Particularly in research fields and small-to-medium enterprises, reduced computing barriers could stimulate more innovative attempts.

With the open-source release of Harness and expanded access to computing resources, the platform is expected to attract numerous developers and research teams to participate in ecosystem development, driving rapid iteration and practical application of agent technologies in real-world scenarios.