en.Wedoany.com Reported - On April 26, China's National Supercomputing Internet platform officially launched a limited-time free dialogue service for DeepSeek-V4. Enterprise users, research institutions, and individual developers can log in to the official Supercomputing Internet website and access the Chat page to experience real-time, fluid dialogues with DeepSeek-V4's million-token ultra-long context for free, eliminating the need to deploy models themselves or bear inference computing costs. Built under the guidance of the Ministry of Science and Technology, the National Supercomputing Internet is a national-level computing power service platform that has now connected dozens of supercomputing and intelligent computing centers. This limited-time free event aims to lower the barrier to using large models and accelerate the penetration of AI technology across various industries.
Released and open-sourced by DeepSeek on April 24, DeepSeek-V4 includes two versions: V4-Pro with a total of 1.6 trillion parameters and V4-Flash with a total of 284 billion parameters. Both natively support a 100 million-token context, utilizing a Mixture of Experts (MoE) architecture and DSA sparse attention technology. On the AIME 2025 mathematical reasoning benchmark and the LiveCodeBench competition-level code evaluation, its performance is on par with GPT-5.4. On the MMLU-Pro general knowledge evaluation, it improved by over 8 percentage points compared to its predecessor. This marks the first time a Chinese self-developed large model has joined the top-tier league of international, most advanced closed-source models in terms of general intelligence. 
The Supercomputing Internet platform is equipped with a dedicated inference acceleration engine for DeepSeek-V4, which leverages KV Cache compression and operator fusion optimization for million-token long sequence scenarios. Under a full load of 1 million token contexts, the initial token latency can be controlled within 0.5 seconds, achieved an inference speed of 60 to 80 tokens per second. The platform also provides a Python SDK and standardized API interfaces, allowing developers to integrate DeepSeek-V4 into their existing business processes with just a few lines of code, enabling long-context applications such as intelligent customer service, legal document analysis, academic paper polishing, and long-form novel continuation.
In terms of computing power resources, this limited-time free service is jointly supported by the National Supercomputing Center in Jinan, the National Supercomputing Center in Shenzhen, and the National Supercomputing Center in Tianjin. Together, these three centers have deployed over 2,000 GPU accelerator cards specifically for DeepSeek-V4 inference services. The first phase of the free event will run from April 26 to May 26. The platform will dynamically adjust resource allocation based on actual call demands. The Supercomputing Internet platform currently has over 120,000 registered enterprise users across more than 20 industries, including advanced manufacturing, biomedicine, financial technology, and advanced materials, has cumulatively completed over 5 billion AI inference tasks.
The Supercomputing Internet has simultaneously integrated DeepSeek-V4 with the platform's existing toolchain call links. After conducting dialogue tests with the model within the same workspace, users can directly use the platform's built-in model fine-tuning, quantization deployment, and stress testing tools to achieve a complete closed-loop process from model selection to application launch. The platform has now launched a total of 32 open-source large models covering multiple domains such as natural language processing, computer vision, and scientific computing. All models support one-click invocation and a unified API specification. Wang Li, Deputy Director of the High and New Technology Development Department of China's Ministry of Science and Technology, previously stated publicly that the Supercomputing Internet aims to become a national-level hub platform for "Model as a Service," accelerating the engineering transformation of artificial intelligence technological achievements by aggregating computing power, models, and the developer ecosystem.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
