Huawei open-sources a robot world model
Huawei Cloud released CloudRobo-WM1.2, its first Ascend-native embodied world model, and said it will open-source the weights, inference and training code, claiming first place on RoboTwin 2.0 and Libero Pro 2.0.
On the afternoon of November 16, Huawei Cloud unveiled CloudRobo-WM1.2, its first Ascend-native embodied world model, at an online launch. The detail worth more attention than the model itself came from Wu Xiaodong, president of Huawei Cloud's embodied products line: the weights, inference code and training code will all be open-sourced.
What kind of model it is
Huawei lists five core technologies: a PPA multimodal world representation architecture, a world-model-native DualAct architecture, a multi-granularity supervised adaptive pre-training framework, body-semantics-preserving cross-embodiment motion modelling, and four composable elastic inference modes.
They map onto three questions. Representation handles how a scene becomes a computable form; DualAct handles generating actions directly on that representation; cross-embodiment handles whether the model still works on a different robot; elastic inference handles graceful degradation when compute is tight. The third matters most — it decides whether the model is tied to one body or can migrate across forms.
Wu's benchmark claim: first place on both RoboTwin 2.0 and Libero Pro 2.0. Both are simulated benchmarks testing manipulation success and generalisation, not the stability of continuous real-robot operation.
First place on a simulated benchmark also deserves a discount. Benchmarks like RoboTwin and Libero have fixed task sets that models can be tuned against, while factory tasks are open: light, materials and human motions change daily. A leaderboard shows a ceiling, not a floor on site.
The weight of "open source"
Releasing weights, inference code and training code together is uncommon among Chinese embodied models. Most vendors open either inference code or a small-parameter version; weights plus the training chain is rare.
The intent is easy to read: world models have no de facto standard yet, and whoever developers adopt first defines the interface. Open-sourcing is the fastest way to spread and the cheapest ecosystem investment — rather than negotiating with every robot maker, let them download and adapt on their own.
The cost is equally clear: once the model is open, others can build products on it, and Huawei captures only cloud and compute usage rather than a premium on the model itself.
Compute is the other backdrop. Over the past two years Ascend's main customers have been internet and government clients training large models, with robotics a small share; if embodied AI scales, it becomes a new and steady source of inference demand. That may be the real motive for giving the model away.
The ecosystem numbers
Platform figures disclosed at the launch: CloudRobo has gathered more than 16,500 professional embodied developers, over 230 companies and universities, and 1,500-plus industry assets. Huawei says 80% of Chinese embodied-AI companies use Ascend Cloud, with key universities essentially covered.
The partner list includes listed companies such as Yijiahe, Tuosda and Fulongma, robot makers such as JAKA, Qianxun, Galbot, Youi and Realman, and on the data side Lingyun Guang's 4D motion capture and spatial reconstruction.
The line used at the launch was that body makers can "arrive with hardware and leave with a model." In commercial terms: Huawei is not competing for the robot sale, but for compute, platform and deployment revenue.
Read alongside OpenAI's reorganisation
Last week OpenAI disbanded its stand-alone robotics team on the argument that general models can control robots directly; this week Huawei released a dedicated world model and open-sourced it. The two look opposite but point at the same thing: the robot's brain is becoming a separate layer rather than a capability inside a robot maker.
The difference is delivery. OpenAI makes general models stronger and lets robot makers connect; Huawei packages model, platform and compute so that makers plug in and use it. The first bets on the ceiling of model capability, the second on the width of an ecosystem entrance.
Huawei will keep releasing models through the first quarter. What to watch is not the leaderboard but three things: whether the open weights are actually integrated into products, how much success rate survives cross-embodiment transfer on real hardware, and whether the "80% of companies" figure turns into revenue next year.
This piece summarizes public reporting and did not independently verify benchmark results or the partner list. Not investment advice.