V4.1 Flash enters China's state supercomputer network

DeepSeek V4.1 Flash became available in the state-run national supercomputing internet centre on launch day, 10 September 2026, the Chinese portal IT之家 reports. Users sign in on the official site, move to the model services section and open the API call interface for V4.1 Flash. So the model is not turning up only at commercial providers. It is going into public infrastructure.

IT之家 also describes the model in terms that match the manufacturer's card: 552 billion parameters in a Mixture-of-Experts architecture, a new Causal-Encoder-Decoder structure, and an input and output asymmetry of 8 billion active parameters on input and 16 billion on output. The portal puts the cost well below comparably sized models. It adds that after reinforcement training at a larger scale, the model beat the flagship DeepSeek V4 Pro and others in benchmark tests.

Memory is what matters most for public infrastructure. According to IT之家, V4.1 Flash cuts HBM demand to one quarter and SSD demand to one eighth of the previous generation. In agent scenarios, cache hits usually make up a large share of the bill, so compressing the KV cache directly lowers the cost of running that kind of task. On shared infrastructure, that is what counts.

Why this is more than a marketing statement

An independent CorCom analysis of 7 September 2026 sets out the background. Its authors at Banque Lombard Odier argue that Chinese players are building an edge not only on model prices but also on energy availability and the pace at which they bring computing capacity online. The American ecosystem keeps its lead in the hardest tasks, cloud infrastructure and proprietary data. Value in the chain, meanwhile, is shifting gradually to semiconductors and large-scale operators.

In that setup, the public supercomputer network is a tool of technology policy. It gives Chinese companies and institutions fast access to the model without intermediaries and without depending on a foreign cloud. DeepSeek is not the only provider here. Competing Chinese models appeared in the country's industry media over the same period, among them the GLM-5.3 family and Kimi from Moonshot. The model's mere availability in that network signals that domestic models are treated as part of critical infrastructure, not just as a commercial product.

For the end user, the difference is practical. A model running in the public network inherits DeepSeek's peak and off-peak pricing, so batch jobs sent on days off cost half of what they cost at times of heaviest load. Low HBM and SSD demand plus the cheaper night rate produces a unit cost that models of this class running only in commercial Western clouds do not reach. It is no accident that in Chinese industry media the competition around V4.1 Flash today is not about benchmark scores. It is about price per million tokens and about capacity availability at peak times.

Read it with some caution. The model's presence in a state network does not mean that all the infrastructure beneath it comes from China. Graphics cards remain largely imported, and Chinese research centres are working hard on variants of domestic accelerators. The public supercomputer network is therefore above all a layer of access and distribution, not proof of full hardware independence.