Models
Huawei Cloud releases CloudRobo-WM1.2, which it calls the first Ascend-native embodied world model, and says it ranks first on two robotics benchmarks; the reports give no scores
Huawei Cloud released and open-sourced CloudRobo-WM1.2, trained on Ascend chips. It says it ranks first on RoboTwin 2.0 and LiberoPro 2.0; no scores given.
Huawei Cloud announced on Monday, November 16, 2026 that it had released and open-sourced CloudRobo-WM1.2, an embodied world model trained on its own Ascend computing platform. A world model predicts how a robot’s surroundings may change after an action, which helps it plan a task. Huawei describes this one as the first Ascend-native model of its kind. The reports we read repeat that claim and a first-place benchmark result, but give no scores, parameter count, repository or license.
The essentials
- What was announced: CloudRobo-WM1.2, an embodied world model trained on Huawei’s Ascend chips, released and declared open source on November 16.
- How it works, per the reports: a shared world-model backbone for predictive world representation, and a ‘DualAct’ architecture that generates actions natively from the world model.
- Benchmarks: reported first on RoboTwin 2.0 and LiberoPro 2.0. Scores and rival models are not given.
- Partners: Chinese press lists listed companies such as Yijiahe, Topstar and Fulongma, and robot makers such as JAKA and Realman, as ecosystem partners.
‘First Ascend-native’ is Huawei’s label
The ‘first’ comes from Huawei Cloud’s own billing of the model. It means the model was trained on Ascend hardware rather than on Nvidia GPUs. One Chinese report adds that rival robot ‘brain’ models such as Gemini Robotics 2.5 and GR00T 2 depend heavily on Nvidia’s ecosystem. That comparison is the reporter’s framing, not a measured result.
Open source, but where?
The reports say the model was open-sourced. The ones we read name no repository, download link or license, so the terms of reuse cannot be checked from them. Until a repository is published, ‘open source’ here rests on Huawei’s statement.
What is not established
- The benchmark scores and which models were beaten: not published in the reports we read.
- Parameter count and compute used for training: not stated.
- Repository and license: not named.
- Whether the model runs on non-Ascend hardware: not stated.
- Any real-robot results beyond benchmarks: not reported.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Release | Released and open-sourced on November 16, 2026 | ChainCatcher |
| Training hardware | Huawei Ascend computing platform | ChainCatcher |
| Architecture | Shared world-model backbone and DualAct native action generation | Futunn |
| Benchmarks | Reported first on RoboTwin 2.0 and LiberoPro 2.0, no scores | ChainCatcher |
| Capabilities | World understanding, future prediction, action generation, scalable learning | TechFlow |
FAQ
What is a world model for robots?
A model that predicts how the environment will change in response to an action, which a robot can use to plan its next move.
Is CloudRobo-WM1.2 really the first Ascend-native world model?
That is Huawei Cloud’s description. We found no independent check of it.
Can I download it?
The reports say it is open source but name no repository or license, so check Huawei Cloud’s official channels.
Method and limits
Checked on November 16, 2026. We read ChainCatcher, TechNode, Futunn and TechFlow through search excerpts and one page fetch, not Huawei Cloud’s own announcement. The request called it a ‘first’ without saying that this is the company’s claim, and did not mention the open-source status.
Sources
- ChainCatcher: Huawei Cloud released and open-sourced the Ascend Native Embodied World Model CloudRobo-WM1.2 (verified 2026-11-08)
- TechNode: Huawei Cloud plans to open-source its Ascend-based world model for robots (verified 2026-11-08)
- Futunn (STAR Market Daily): 华为云发布并开源昇腾原生具身世界模型 (verified 2026-11-08)
- TechFlow: Huawei Cloud Releases and Open-Sources Embodied World Model CloudRobo-WM1.2 (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy