China's National Data Administration released a draft action plan calling for a batch of industry high quality datasets covering key sectors and validated through application by the end of 2028, Jiemian reported. The plan asks regulators to develop data driven AI applications, build specialized data companies and talent, and codify standards for dataset construction. China has already built more than 100,000 high quality datasets nationwide by the end of 2025, and more than 30 new standards covering public data, data infrastructure and AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. are expected to be issued this year. The administration was set up in 2023 to serve as Beijing's central data regulator. The directive puts the National Data Administration at the center of China's data and AI governance stack, formalizing a state led pipeline for training data supply that has no direct U.S. counterpart.
Read at Jiemian ↗
Alibaba upgraded its large language model organizational structure on Monday, merging the Tongyi large model team with the Future Life Lab to form a new Token Foundry business unit reporting directly to CEO Wu Yongming, Jiemian reported. The reorganization names Zhou Jingren as chief scientist and stands up an AI Future Research Institute. The move builds on Alibaba's March 2026 creation of the Alibaba Token Hub group, also placed under Wu's direct control, treating token output and distribution as the company's core commercial AI layer. Alibaba's Qwen models sit inside this restructured stack.
Read at Jiemian ↗
A consortium led by Shenzhen Hetao College, with Harbin Institute of Technology Shenzhen, the Shenzhen Big Data Research Institute and Huawei, completed full-parameter post-training of the 1.6 trillion-parameter DeepSeek-V4-Pro model on a Huawei Ascend 910C cluster, Pandaily reported. Until this run, domestic accelerators were viewed inside the industry as suitable for inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. but not for end-to-end training at frontier scale. The Mixture-of-Experts training workload requires every expert module to update in parallel, producing all-to-all communication volumes dozens of times higher than standard dense models. The achievement narrows the strategic premise that Bureau of Industry and Security controls on Nvidia's H100 and H200 chips cleanly bottleneck Chinese frontier model development, a premise underpinning current export license design.
Read at Pandaily ↗
Chinese AI models have led American models in total weekly API call volume on OpenRouter for six consecutive weeks, with Chinese models drawing 14.19 trillion tokens against 3.2 trillion for U.S. models in the most recent week tracked, Pandaily reported. DeepSeek-V4-Flash held the top spot and Tencent's Hy3 preview took second, while Shanghai-based MiniMax M3 broke into the global top three for the first time and Xiaomi's MiMo-V2.5 entered the rankings at fourth. The report credited aggressive pricing and open-weightOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. distribution for the developer adoption gap. OpenRouter is one of the largest AI model routing platforms, providing a single, comparable usage signal across vendors. The sustained lead sharpens the U.S. policy debate over whether export controls and any forthcoming federal rules on open-weight models should target inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. reach and downstream deployment, not chips alone.
Read at Pandaily ↗
Moonshot AI, the developer of the Kimi assistant, opened a new fundraising round at a pre-money valuation of $30 billion, up from about $4.3 billion at the end of December 2025, TechNode reported, citing National Business Daily. The round follows a previous financing that closed at roughly $20 billion post-money in early May. Meituan Longzhu partner Wang Xinyu said Kimi's annual recurring revenue surpassed $200 million in April after the release of the K2.5 model. The deal lands Moonshot in a four-company front rank of Chinese model developers alongside DeepSeek, Zhipu and MiniMax. The compressed funding cadence shows private capital pricing in continued growth at Chinese model labs even as the Commerce Department tightens enforcement of Nvidia and AMD advanced chip license requirements on offshore affiliates of Chinese firms.
Read at TechNode ↗
JD.com and Tencent have agreed to collaborate on AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise., pairing JD's e-commerce and fulfillment systems with Tencent's WeChat user network, TechNode reported, citing IThome. JD's agent has been wired into native AI assistants on Huawei, Oppo and Honor devices through an Agent-to-Agent protocol, letting users issue shopping requests that JD's logistics and customer service network then fulfills. Tencent is separately working with Xiaomi and Vivo on similar A2A integrations for WeChat. The setup gives Tencent's WeChat a gatekeeper role over how third-party AI agents call into handset level assistants. Concentrating agent traffic through WeChat and JD's commerce stack hands the Cyberspace Administration of China, already the licensing authority for generative AI services, a narrower regulatory chokepoint as agent-to-agent activity scales inside Chinese consumer apps.
Read at TechNode ↗