A Ministry of Commerce spokesperson yesterday rejected the U.S. allegation that Chinese AI companies carried out industrial-scale distillation of American models, calling it groundless. Routing the answer through the trade ministry, which handles export controls, frames any resulting curb on model access as a trade measure rather than a security one. The advisory, issued Tuesday by the National Security Agency, the Federal Bureau of Investigation and the Cybersecurity and Infrastructure Security Agency, recommended defensive steps to American firms. Distillation trains one model on the outputs of another, transferring behavior without access to the original weights. The spokesperson called it a neutral technique that U.S. labs also apply to Chinese models, and said Washington's use of anti-distillation measures to suppress Chinese companies would draw countermeasures. China is willing to join an intergovernmental dialogue on AI that the two countries' leaders have agreed to hold, the spokesperson said.

Amap released ABot-Earth 0.7, which the map maker calls the world's first 3D-native city world model. A world model learns how a space is laid out and how it changes. Amap's is trained end to end on 3D data rather than fitted to flat imagery as Google's pipeline does. The training data is 3D Gaussian splatting, a format that represents a scene as millions of colored 3D blobs. Amap says the model turns one satellite image or a text description into a kilometer-scale 3D city scene in about 10 minutes on a consumer graphics card. Output exports into game engines including Unity and Unreal Engine, and the company puts 3D mapping cost at roughly 1% of the traditional method's. Amap says it trained its Gaode Tutu robot inside an ABot-Earth simulation and, in Changzhou, Jiangsu, has tied ground vehicles, low-altitude drones and robots into one city-scale unmanned system.

Zhiyuan Robotics, which uses the English name AgiBot, released GE-Act 2.0, a control model trained from scratch on robot manipulation data rather than built on an existing video generation model. If capability tracks data volume, Chinese robot makers can buy it with deployment scale rather than research breakthroughs. The model predicts both how a scene will change and what the robot should do next, so perception and control sit in one network. Evaluated on real robots with no fine-tuning for the test tasks, the G1-OP machine's success rate rose from 17.1% to 44.1% as training data grew a hundredfold. A second robot body that supplied only a sliver of that data gained 17.7 percentage points, which AgiBot reads as skill transferring between different machines. After fine-tuning it beat the pi0.5 and GR00T N1.7 models on benchmarks for unfamiliar environments and instruction following, the company says.

JD.com released JoyAI-Echo WM, a real-time interactive world model it says scored 81.6 on the WBench navigation benchmark, first among interactive world models. The release came at the same JDDiscovery conference in Beijing as the 100,000-GPU cluster covered on September 9th. Publishing the data rather than only the model puts JD's own record of how people do physical tasks into other labs' training runs. The dataset, EgoLive, is the first release from 10 million hours of human embodied-task recordings JD Cloud is collecting, and universities and research institutes have asked to use it. In logistics, the company says its Super Brain 3.0 model cut end-to-end route solving for hundreds of millions of packages from minutes to seconds. Over the next five years JD says it will buy 3 million robots along with unmanned ground vehicles and drones.