On September 17, China Telecom Artificial Intelligence Technology Co. officially released Xing4.0 (Xingchen / 'Stars'), a new-generation lightweight agent model purpose-built for the agent era.
The model adopts a Mixture-of-Experts architecture with 29 billion total parameters but only 4 billion activated — making it the first large model in China trained entirely on domestic compute and a domestic framework and deeply optimized for complex engineering tasks.
Xing4.0 focuses on complex task understanding, task planning, tool calling and autonomous execution, with long-context handling. It is designed for code development, data analysis, office automation and lifestyle services, positioning it as a practical 'digital employee' rather than a chat tool.
The release reflects a broader shift in China's AI industry from capability demonstration toward industry-wide deployment. With the core AI sector already exceeding 1 trillion yuan in scale under the national 'AI+' action plan, telecom and cloud operators are racing to ship agent-ready models that can be embedded directly into enterprise workflows.
By training on domestic hardware and software stacks, China Telecom signals confidence that a fully localized pipeline can produce competitive large models — a strategic hedge as supply constraints on advanced overseas chips persist.
