A launch slide puts Huawei Cloud's CloudRobo-WM1.2 at 88.98% average on RoboTwin 2.0. Second place: 86.47%.
The same model takes 81.7% on LIBERO-Pro 2.0, where second is 75.0%. Huawei Cloud published both on November 16.
Two first places, both self-reported.
A 2.5-point gap is a 19% gap in failures
88.98% success means 11.02% failure; 86.47% means 13.53%. Divide and you get 0.814 — 18.6% fewer failures on the same tasks.
LIBERO-Pro 2.0 widens it: 18.3% against 25.0%, or 26.8% fewer failures.
A 2.5-point leaderboard gap is nearly a fifth of the failures. Manipulation is graded on the tail: a robot is judged on whether it fails the same motion repeatedly.
Both numbers appear only in Huawei Cloud's launch material, and no third party has reproduced them.
Open source, not yet downloadable
The launch was clear: model, training framework and Ascend compute practices, all opened together.
CloudRobo's model assets list CR-VLA-Dual, Lerobot-ACT and Lerobot-π0.5, with no WM1.2 weights.
Reporters who attended wrote the same: no repository address on launch day, no licence, no parameter count. Without a licence, commercial use and redistribution have no clear grant.
Open source and downloadable are different claims: the first is a promise, the second is code that runs.
Huawei Cloud says the framework and Ascend practices will also be released. When the 1.2 weights arrive, how much, under what licence, the launch did not say. "Open source" cannot yet stand in for "usable".
Guessing words, guessing frames
A language model has one objective: given preceding text, guess the next token. The vocabulary is discrete, the loss is cross-entropy, and a wrong word can often be recovered from later context.
A world model answers: given the current image and an action, what does the scene look like one step later?
Input is RGB, depth, point clouds, joint angles and torque; output is the next frame, state and motion. Loss lands on pixels, latents or 3D representations. Four dimensions: three of space, one of time.
A language model can lose a word and the reader may not notice. A world model that loses a frame puts the robot into a wall.
Worse, error compounds. At 99% accuracy per step, the chance of 100 correct steps is 0.99 to the hundredth power, 36.6%. Real systems close the loop, so it will not be that bad. The order of magnitude still holds.
One second of video, 36,000 tokens
A 640 by 480 video at 30 frames per second, cut into 16 by 16 patches, gives 1,200 tokens a frame, 36,000 a second. A 2,500-character article is roughly 2,000 to 3,000 tokens. One second equals a dozen articles.
A language model scrapes text. A world model collects on the line — teleoperation, trial and error, simulation — every route dearer than crawling a page.
Expensive data, heavy tokens, compounding error: that is why computation runs in latent space, not on pixels, with actions split out.
DualAct splits the two paths
CloudRobo-WM1.2 shares one backbone, then forks.
VideoDiT predicts future video representations for accuracy. ActionDiT is an action expert emitting motion directly for speed. Both train together; at deployment the system can switch to the expert to cut latency.
Four inference modes: World, Expert Action, Joint, Action-only. The names matter less than the admission: a world model is accurate but slow, and a robot needs both. The price is training complexity: branches must stay aligned, and the expert's capacity needs separate tuning.
PPA handles input, folding RGB, semantics, depth and point cloud into one representation before predicting a multimodal future. The launch calls it moving from reading a still photograph to reading a 4D action in progress. Marketing, but the target is real: not recognising objects, but anticipating what an interaction produces.
Counting cooling towers from orbit
Two Epoch AI researchers, Ben Cottier and Lynette Bye, published "Huawei's Horinger data center reveals China's compute gap." They used commercial satellite imagery, not Huawei's press release.
The method is simple: count towers. The campus has west, south and north zones. The north has visibly fewer towers, so the researchers excluded it as non-AI space. West and south hold four buildings with high-density cooling; from tower type, size and count they estimated heat rejection, backed out IT power, and reached 286 megawatts.
Cross-check: a nearby substation is rated around 304 MW, backup generators around 230 MW. The 286 sits between them.
Then power became chips. Using Huawei's sales records and a four-month deployment lag, they estimated about 105,000 Ascend 910C and 8,000 910B running, three of four buildings live. Normalising — Huawei discloses 910C at about 0.76 of an H100, Nvidia's B200 and B300 at about 2.5 — Horinger comes to roughly 81,000 H100-equivalents, with a 90% confidence interval of 36,000 to 141,000.
81,000 is not 81,000 cards
81,000 is H100-equivalent compute, not cards. The physical inventory is 105,000 910C plus 8,000 910B, or 113,000 chips. And 286 MW covers four buildings while 81,000 describes three; dividing them gives an artificially low compute-per-megawatt, since the denominator includes a hall not yet switched on.
On Epoch's basis, Colossus 2 is about 1.14 million H100-equivalents, exactly 14 times 81,000. The ratio is 7.1%.
By 2028, Horinger is projected at 141,000, up 74%; Colossus 2 at least 1.8 million, up 58%. The ratio becomes 7.8%.
Growing faster, holding the same share: the ratio barely moves.
The 90% interval spans 36,000 to 141,000, a spread of 3.9 times. This is a third-party estimate, not a Huawei disclosure.
On one side, 88.98% on a slide. On the other, 81,000 from orbit.
One is capability, the other a compute base. Neither has been independently checked. What can be verified is the method — one tests a model on simulated tasks, the other counts cooling towers.
Methods outlast conclusions.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
