On November 16, 2026, Huawei Cloud officially unveiled CloudRobo-WM1.2 at an online launch event, the first Ascend-native embodied world model on its Industry AI Dream Factory platform, and announced it will be open-sourced. The model ranks first on both robot manipulation benchmarks RoboTwin 2.0 and LiberoPro 2.0, with average scores of 88.98% and 81.7% respectively.
What exactly is a world model computing for a robot
A world model's core job is to predict, before the robot moves, what will happen to its surroundings if it takes a given action. CloudRobo-WM1.2 introduces five innovations, two of which matter most: the PPA multimodal world representation architecture uses a shared world-model backbone to produce prediction-enhanced world representations, giving the model a more complete understanding of the scene; the world-model-native DualAct architecture lets actions be generated directly by the world model, instead of planning first and handing execution to a separate module. The model was trained entirely on domestic Ascend computing power. The event was hosted by Wu Xiaodong, president of Huawei Cloud's embodied product line, with chief model expert Liu Liu presenting the architecture. Huawei Cloud said at the event that 80% of China's embodied AI companies have already chosen Ascend Cloud.
Two things the launch did not make clear
First, the exact form of the open source release. The event announced open sourcing, but as of writing, no weight download address or code repository has been published, so developers still have to wait for Huawei Cloud to release the details, and it is unclear when they will actually be able to try it. Second, the benchmark figures. RoboTwin 2.0 and LiberoPro 2.0 are both public manipulation benchmarks, but the two first-place finishes currently appear only in Huawei Cloud's own launch materials and have not been independently verified, so the scores still need to be confirmed by later hands-on testing.
For developers, the real point of this launch is not launch day itself but what happens after the open-source repository goes live: whether the weights are complete, whether it can run at low cost on Ascend hardware, and how far on-device inference latency can be pushed — those are what will decide whether it gets used in real projects. Runway's Praxis-1 is another open world-action model worth comparing, to see how the two approaches differ: Runway Open-Sources Praxis-1: a 27B World-Action Model Tops RoboLab-120, Ready for Robots to Use.