Huawei Cloud has released CloudRobo-WM1.2, an open-source world model for robots trained on its Ascend chips. The announcement moves beyond last week’s preview: Huawei says the model generates robot actions directly and ranked first on two named benchmarks.
Huawei AI Watch analysis
What happened
Huawei Cloud unveiled the model at an online event, according to AI Insider’s roundup of developments in China’s physical-AI sector. The company describes CloudRobo-WM1.2 as using a DualAct design that generates robot actions directly from the world model. Huawei also says it is releasing the training framework and recommended methods for running the software on Ascend systems.
The company claims first place on the RoboTwin 2.0 and LiberoPro 2.0 benchmarks. The report gives no scores or independent comparison, so the ranking is a Huawei claim, not a settled verdict on how the model performs across real robots and tasks.
Why it matters
An open model plus its training framework could give robotics developers something more useful than a launch-day demonstration: a starting point they can inspect, adapt and run within Huawei’s AI-computing ecosystem. That makes the release relevant both to robot builders and to the wider question of whether China’s domestic AI stack can support practical robotics development.
The useful test is not the benchmark podium alone. It is whether outside developers can reproduce the results, use the framework on supported Ascend systems and get reliable behaviour on physical hardware.
Our read
This is a substantive release, not just another date on an event poster. The open-source model and accompanying framework are the concrete news; the benchmark lead is promising, with the usual asterisk that the claim comes from Huawei. Developers should look for the released code, deployment details and reproducible tests before treating “first” as the end of the conversation.
What to watch
- What model, training code and documentation Huawei makes available.
- Whether independent teams reproduce the benchmark results.
- Which robots and Ascend systems can run the model, and how it performs outside benchmarks.
Discussion spark: For robot world models, should developers put more weight on open code they can test themselves, or benchmark results claimed by the company that built the system?
Sources and evidence
- China Physical AI Roundup: UniX AI Raises $75M Pre-A, Plus TensorWorld, UBTech Order and Huawei CloudRobo – AI Insider (17 November 2026, 10:05 UTC)
Independent WittyWires tracker for public updates about Huawei AI. Not affiliated with or endorsed by Huawei AI; this is not an official account.