Huawei Cloud open-sources its CloudRobo-WM1.2 robot world model, and the price of entry is Ascend compute
Huawei Cloud released an embodied world model trained on its own Ascend chips and said it will open the weights and code. The model is the free part: the training framework and Ascend best practices that ship with it are built to keep China's robot makers training on Huawei's cloud.
Huawei Cloud released CloudRobo-WM1.2, an embodied world model trained on Huawei's own Ascend chips, at an online launch on Nov. 16, and said it will open-source its weights, inference stack and code, FX168 reported. The model predicts how a scene will change and generates robot actions from one shared backbone, through what Huawei calls a DualAct architecture, according to IT Home. Huawei says it ranked first on two simulation benchmarks, RoboTwin 2.0 and LiberoPro 2.0.
The model is the free part; the compute underneath it is the product. Huawei Cloud says it does not build robot bodies and leaves them to partners. Along with the model it is opening its training framework and its "WAM" best practices for Ascend Cloud, IT Home reported, and it says 80% of Chinese embodied AI companies already choose Ascend Cloud. An open model tuned for Ascend gives the other 20% a reason to train there too, at a time when Chinese investors are already moving robot money into domestic chips.
The partner list reads like a customer list
Huawei launched the CloudRobo platform in June 2025. It now counts more than 16,500 embodied-AI developers and over 230 companies and universities, the company says, according to Sina. Its partners include listed companies Yijiahe, Topstar and Fulongma and unlisted robot makers such as JAKA Robotics, Galaxea and Youibot. Huawei's pitch to hardware makers, as Sina quoted it, is to "come in with hardware and leave with a model".
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.
Some of that work is already in the field. Topstar has deployed a wheeled humanoid with Huawei Cloud for material handling in factories, from scanning bins to palletizing, and in September Huawei Cloud and Yijiahe launched the Z200Pro, a robot for live-line work on power distribution networks, Shanghai Securities News reported.
The strongest point against the lock-in reading is the license: open weights can in principle be ported to other chips, including NVIDIA's. Huawei's bet is that a model built "Ascend-native", and a training framework to match, make its cloud the cheapest place to keep improving it.
The launch closed with a street sweeper. A new smart cleaning robot developed with sanitation-equipment maker Fulongma is on presale, with a launch in Shanghai in December, The Paper reported. Huawei Cloud says its next model will follow in the first quarter of next year.
Sources Sources verified
- 1.The Paper, Nov 16, 2026: 华为云“行业AI梦工厂”发布并开源昇腾原生具身世界模型thepaper.cn
- 2.FX168, Nov 16, 2026: 华为云发布具身世界模型,物理AI再迎重磅催化!fx168news.com
- 3.Sina, Nov 16, 2026: 华为昇腾云在具身智能行业市场格局中国第一k.sina.cn
- 4.IT Home, Nov 16, 2026: 华为云发布并开源昇腾原生具身世界模型 CloudRobo-WM1.2ithome.com
- 5.Huawei Cloud, Nov 16, 2026: 华为云CloudRobo具身世界模型发布会huaweicloud.com
- 6.Shanghai Securities News via 21Jingji, Nov 12, 2026: 华为云宣布,首款昇腾原生具身世界模型来了m.21jingji.com
- #world model
- #Huawei Cloud
- #Ascend
- #open source
- #China
Share this briefing
Know someone who follows robotics? Send it their way.
Related briefings
More in Embodied AI & Foundation ModelsAgiBot's GO-3 Preview model took a store robot from 28.6% to 85.7% success in five days, and opens up at year-end
AgiBot unveiled GO-3 Preview, a robot foundation model it plans to open-source by the end of 2026, and says a G2 running it in a Shanghai convenience store went from 28.6% to 85.7% full-task success in five days. AgiBot is selling the model on how fast it learns on the job; both store figures are sevenths, so the gain may rest on a few dozen runs.
Hesai's Kosmo turns real rooms into robot training worlds in minutes
The lidar maker launched Kosmo One (8K spatial camera) plus Kosmo World cloud, partnering with Galbot and Dexmal; a Dairy Queen scan cut modeling from ~300 hours to 15 minutes.
Skild's S1 learns soccer by playing itself for 140 simulated years
After launching S1 — a robotics foundation model that copies a single video demo without fine-tuning — Skild showed Physical Self-Play: the same base policy trained against itself in Isaac Sim, then transferred to a humanoid.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.