A world model that acts, not just predicts

Longbridge's event report describes joint modeling of perception, prediction and action, with a dual-path action design. For complex tasks the model predicts the future scene and the action together. For simple, high-frequency operations it can skip part of the prediction and output the action directly.

That trade-off is the practical core of the design. Prediction costs compute, and latency requirements depend on the control task. The report says Huawei Cloud offers four composable inference modes whose compute requirements differ by roughly a factor of ten. Liu Liu, the Embodied Intelligence Lab's chief model expert, put current inference latency at about 90 milliseconds, with the next stage aiming for under 60 — a figure the reports present without the hardware, workload or measurement definition needed to generalize it.

The benchmark claim has no numbers attached

Huawei Cloud said the model achieved first place on RoboTwin 2.0 and LIBERO Pro 2.0. Those are its own disclosed test results, and the launch reports carried no scores, no margins and no list of the models it beat. Until the weights and code arrive, there is no way for an outside team to reproduce either ranking.

The training figures need the same care. Liu said the test results were previously achieved on about 8,000 hours of pre-training data, and that the current training scale has expanded to about 27,500 hours from 20 data sources covering 33 types of robot embodiment. The reporting does not clearly identify which checkpoint earned the first-place rankings, and that is the question that decides how much weight the rankings can carry.

"Open-sourced" on November 16, downloadable on November 30

The launch coverage described the model as released and open-sourced on November 16. The announced plan in the same event reporting is narrower: on November 30, Huawei Cloud will open the weights, the inference code and the post-training code, and will selectively share pre-training experience and related capabilities with certain partners.

Two details follow from that. First, the reports schedule public weights and code for November 30, and ALFW did not locate first-party downloadable artifacts or license terms — so for the two weeks between launch and release, every capability claim rests on Huawei's account as relayed by the launch coverage. Second, "open-source" currently describes an intention rather than terms anyone can read. A November 30 release could permit outside testing, but reproducing the rankings would also require the evaluated checkpoint, the harness and the settings — weights alone would not validate every launch claim.

Why Huawei is building the brain as well as the cloud

Wu Xiaodong, who leads Huawei Cloud's embodied product line and its Embodied Intelligence Lab, was asked at the launch whether Huawei Cloud would become a direct competitor to the model companies that build on its platform. His answer was no: the stated purpose is to publish training experience and engineering practice on Ascend so partners can train their own models, and he said Huawei Cloud would, on equal conditions, prioritize recommending partners' models in industry solutions.

ALFW's reading: a public reference implementation could help partners evaluate development on Ascend, which is reason enough for a cloud provider to publish one. The ecosystem figures given at the event — the Embodied Intelligence Zone, launched in June 2026, has drawn more than 230 enterprises and universities, more than 16,500 developers and over 1,500 digital assets — describe the audience Huawei Cloud is building for. The Fulongma sanitation robot shown at the same event is in pre-sale ahead of a December 17 release in Shanghai; the reporting does not establish that it runs CloudRobo-WM1.2, and it is event context, not evidence about the model.

Sources

  1. [1] Longbridge, "Huawei Cloud Releases Embodied World Model, Responds to…" (detailed November 16, 2026 launch-event report; named executives Wu Xiaodong and Liu Liu; read first-hand by ALFW)Read source
  2. [2] TechNode, "Huawei Cloud plans to open-source its Ascend-based world model for robots", November 16, 2026 (read first-hand by ALFW)Read source
  3. [3] STAR Market Daily report via Binance Square (November 16, 2026; first-place claims)Read source