Reportedly: Beijing Academy of Artificial Intelligence and Qingmao Intelligence launch commercial FlagOS token factory on Chinese-made chips
Reported by Leiphone · not yet confirmed by the company or a second independent outlet. We update this page when it is.
The Beijing Academy of Artificial Intelligence and Qingmao Intelligence launched a commercial FlagOS token factory on Chinese-made chips, Leiphone reported. The partners unveiled an end-to-end inference optimization solution at Qingmao Intelligence's partner conference on November 17, 2026.
China context
- Original name
- 北京智源人工智能研究院
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- FlagOS is described as open source, but that does not establish access to the commercial service. Before planning an integration, verify whether overseas accounts are eligible and whether the service supports the required model.
Translated from Chinese. Quotes and facts link to the original sources.
Leiphone reported that the Beijing Academy of Artificial Intelligence and Qingmao Intelligence jointly announced an end-to-end inference optimization solution and brought a commercial token factory online. The deployment uses Chinese-made chips and FlagOS, an open-source AI system software stack designed to support multiple chip types.
The approach uses a common software stack so teams do not have to start low-level adaptation and operator rewrites from scratch for each model and chip combination. The report does not identify which chips and models the commercial deployment supports or quantify its performance. A compatibility matrix and throughput and latency results under a stated workload would help verify what the stack delivers.
For operators of Chinese-made accelerator clusters, the reported deployment provides a commercial example of using a shared inference stack instead of maintaining a separate stack for each chip type. The report does not establish how much integration work it saves. Deployment documentation or a customer account showing how an existing model was moved between chip types would help test that benefit.
The commercial test is whether the deployment can deliver a lower cost per million tokens at comparable latency and reliability. The report provides no pricing or operating cost data, so it does not establish a cost advantage.
The next verifiable signals are a supported chip and model list, published service terms, and a named customer deployment with measured throughput and latency. The report does not establish whether customers outside China can use the service; published account eligibility or regional access terms would answer that question.