NextFin News — The National Supercomputing Internet synchronized the official DeepSeek-V4-Flash API and model download service on August 3 as DeepSeek opened the formal version to public beta.
Users can access the API through the platform’s official website by selecting the model-service section under the homepage services menu, or download the model weights from the Supercomputing Internet AI community and complete fine-tuning, deployment and adaptation inside a trusted Notebook environment. The DeepSeek-V4-Flash-0731 release retains the same mixture-of-experts architecture of 284 billion total parameters with 13 billion activated per token and a one-million-token context window that characterized the earlier preview, yet extensive post-training has substantially strengthened agent capabilities and instruction following. Benchmark scores on agent-oriented suites, including Terminal Bench 2.1 at 82.7, now surpass those of the larger DeepSeek-V4-Pro preview while API pricing remains markedly lower than competing closed-source offerings.
The National Supercomputing Internet, China’s first 100,000-card-scale supercomputing-AI fusion resource pool, already hosts more than 1,700 domestic and international open-source large models. By integrating the newly stabilized DeepSeek-V4-Flash endpoint, the platform enables enterprises and developers to invoke the model without local infrastructure configuration, accelerating practical deployment of long-context and agentic applications across research and commercial workloads.
Explore more exclusive insights at nextfin.ai.
