Huawei Cloud Open-Sources a Robot World Model Trained on Ascend Chips

CloudRobo-WM1.2 ranked first on two robot-manipulation benchmarks, and Huawei says 80% of China’s embodied AI companies already use Ascend cloud.

WHY IT MATTERS

Robot makers get a free, competitive model tuned for Huawei's Ascend cloud, which gives them a reason to train there instead of on Nvidia-based clusters.

Huawei Cloud released CloudRobo-WM1.2, a world model for robots trained on Huawei’s Ascend chips, at an online launch on November 16 and said it would make the model open source. A world model lets a robot predict how its surroundings will change before it moves, and then pick the action. IT Home reported that the model ranked first on two public robot benchmarks, RoboTwin 2.0 and LiberoPro 2.0, which test simulated arms on manipulation tasks.

The release is Huawei’s bid to make Ascend the default training platform for China’s robot makers. Huawei Cloud has said since it launched its CloudRobo robotics platform in June 2025 that it will not build robots and will leave the hardware to partners, TMTPost’s English edition reported at the time. A detailed account by TMTPost notes that the full training pipelines behind leading world models had previously run only on Nvidia’s CUDA software or Google’s TPU chips, which makes an Ascend-trained model that tops these rankings a public test of domestic computing.

What has actually been published is less clear. Reports differ on whether the weights are already downloadable or arrive later this month, and none of the reports reviewed named a license or a repository.

The model leads two benchmarks and was built for Ascend from the ground up

Huawei Cloud’s event page set aside a slot in the 90-minute livestream for the "official release and open-sourcing" of the model. Wu Xiaodong, president of Huawei Cloud’s embodied product line and director of its Embodied Intelligence Lab, presented it. The model is the first world model native to Ascend from Huawei Cloud’s Industry AI Dream Factory (行业AI梦工厂), an industry AI platform.

Kuai Keji, in a report republished by Sina, gave average success rates of 88.98% on RoboTwin 2.0 and 81.7% on LiberoPro 2.0. The model shares one backbone between predicting the scene and generating actions, an architecture Huawei calls DualAct. A cross-embodiment module acts as a translator between single-arm, dual-arm and wheeled robots, so that data from different machines can train one model. Its pretraining also accepts incomplete data, such as clips missing action labels, depth readings or camera angles.

Developers can choose among four inference modes to trade accuracy against cost. TMTPost put the range at 3.99 to 43.71 trillion floating-point operations per inference, a gap of nearly 11 times. It also reported that RLinf v0.4, a reinforcement learning framework, now achieves up to 2.78 times higher throughput on Ascend processors.

The release is a sales tool for Ascend cloud, and the terms differ by audience

Huawei frames the model as an on-ramp. It says 80% of Chinese embodied AI companies already use Ascend cloud and that key universities are almost fully covered; those are Huawei’s figures and no outlet verified them. A Caiwen report adds that the CloudRobo development platform, launched in June 2025, has more than 16,500 developers, over 230 companies and universities, and more than 1,500 industry assets. Kuai Keji reported that approved projects can receive up to RMB1 million in Ascend cloud computing support.

Wu said the model is meant to help partners, not compete with them. "We hope to contribute the best implementation on Ascend so that more partners and customers can train their models well on Ascend cloud," he said, according to Caiwen. He also argued that no single company can master many industries alone.

Not everything is open to everyone, or at once. The weights, inference code and post-training code open to developers on November 30, according to Caiwen, which quoted Wu’s announcement; pretraining best practices go only to selected customers and partners. According to TMTPost, the CloudRobo platform opens its interfaces and plugins to third parties, while the training framework’s engineering code and the Ascend computing library for world action models go only to core partners.

Data costs and the next releases will decide how much developers use it

The economics behind the push are in the data. Caiwen reported that collecting data on real robots costs RMB500 to RMB2,000 an hour, which is why Huawei stresses a model that can learn from mixed and imperfect data. The agenda also included a preview of a product developed with Fulongma (福龙马), and Caiwen listed industrial, power, sanitation and healthcare settings among CloudRobo’s uses.

Chinese founders working on world models have cautioned that the technology is still far from factory floors, and benchmark wins in simulation do not settle that question. TechNode described the release as bringing a robotics world model to developers on Huawei’s own hardware.

Huawei Cloud plans further models, including vision-language-action and navigation models, and says new releases will come in the first quarter of next year. The nearer marker is November 30: whether the weights and code appear in a public repository by then, and under what license, will show how open the model really is.