Recently, DeepSeek's Deepseek-R1, V3, Coder and other large models officially access the national supercomputing Internet platform. According to reports, DeepSeek-R1 focuses on natural language processing, with strong text generation capabilities; DeepSeek-V3 focuses on image generation and can generate high-resolution and detailed images, suitable for multiple scenarios such as advertising and art creation. DeepSeek-Coder is a code generation model that helps developers build applications quickly and efficiently.

The supercomputing Internet operates a supercomputing center based on Internet thinking, and connects the capabilities and resources of computing power supply, application development, operation services, users and other parties in the industrial ecosystem to build an integrated supercomputing power network and service platform. Its goal is to closely connect the supply and demand sides, through the market-oriented operation and service system, to achieve the overall scheduling of computing resources, reduce the threshold of supercomputing application, and drive the development of computing technology to a higher level, promote the in-depth application of independent core software and hardware technology, and radiation to drive the development and maturity of independent controllable industrial ecology.

In order to promote the integrated operation of supercomputing power and build a national computing base, the national supercomputing Internet platform was officially launched in April last year, covering various fields such as scientific computing, industrial simulation, and artificial intelligence (AI) model training.

The reporter learned from the national supercomputing Internet platform that as of January this year, there are more than 6,500 types of computing power products on the platform, including nearly 240 model services, including domestic open source large models such as DeepSeek and Qwen. There are also Llama, Stable Diffusion, Gemma, Mistral, StarCoder, ChatGLM and other foreign open source models.

For the launch of DeepSeek, the relevant person in charge of the national supercomputing Internet platform said that the integration of advanced domestic AI technology and the national comprehensive computing service platform is a cutting-edge practice to accelerate the ecological development and popularization of domestic AI technology, which will promote the ecological innovation and achievements of AI technology in terms of market, technology, innovation, products and users. Promote the diversification of AI+, and promote the application and industrialization process of China and even the global AI market. Through the national supercomputing Internet platform, users do not need to download local, online deployment, reasoning, customized training and development of DeepSeek and other well-known open source large models at home and abroad, reducing the threshold for the use of AI technology. At the same time, the universality of computing services has also been improved, and the national supercomputing Internet platform relies on the domestic deep computing intelligent acceleration card and the national integrated computing service system to provide rich heterogeneous computing resources for models such as DeepSeek and provide continuous computing power guarantee for domestic AI application and innovation.

The DeepSeek series models are on par with the world's top big models in terms of performance, but have a significant cost advantage. The DeepSeek series significantly reduces the consumption of computing resources through algorithm optimization, and combines the ability of the national supercomputing Internet platform to be "inclusive and easy to use, rich and easy to use", so that more enterprises and users can afford the application of AI technology and accelerate the popularization of AI technology.

The industry generally believes that the large-scale landing of the DeepSeek model in the supercomputing Internet platform will change the status quo of the domestic AI industry relying on overseas closed-source models. (Reporter Yang Xue)