Moke ворвалась в топ WorldArena: мировая модель робота на 32 чипах Zhenwu 810E

Chinese company Moke has unveiled a world model for robots that has shot to the top of the WorldArena benchmark after training on just 32 Zhenwu810E accelerators — according to a report published on July 30, 2026 by the outlet 机器之心 (Jiqizhixin).

What problem does Moke solve

Today's robot world models struggle to hold onto the physics of the real world: a gripper fails to pick up an object, a robotic arm doesn't interact properly with its environment, and once objects in the frame become occluded, the prediction falls apart entirely. The problem isn't "picture quality" — AI-generated video of the future can look crisp frame by frame, yet diverges from reality the moment contact occurs. This is precisely the flaw Moke's model targets: it aims for a physically accurate forecast of what will happen when a robot interacts with objects. According to 机器之心 (Jiqizhixin), current world models repeatedly fail on exactly these scenarios — grasping, contact, occlusion.

"On screen, the simulated robotic arm moves along a perfect

trajectory, but in the real world the cup is still sitting on the table — or has already been knocked over by the gripper," — from the 机器之心 (Jiqizhixin) report.

Key facts of the news:

• Developer — the Chinese team Moke (莫刻), previously a little-known "dark horse"
• Model — a world model for robots that climbed to the top of the WorldArena leaderboard
• Hardware — just 32 Zhenwu810E (真武810E) accelerators
• Publication date — July 30, 2026, source 机器之心 (Jiqizhixin)
• Task — accurate prediction of grasping, interaction with the environment, and scenes involving object occlusion

Why just 32 chips matters

The main twist in this story is computational efficiency. Moke reached the top of WorldArena using only 32 Zhenwu810E accelerators, whereas climbing to the top of similar leaderboards usually requires noticeably larger clusters. For the industry, this is a signal: a competitive world model can be trained on a compact hardware fleet rather than thousands of chips. Zhenwu810E is a Chinese AI accelerator, and betting specifically on it underscores a push toward technological independence from imported GPUs. The number "32" itself is highlighted in the report's headline as a record: that many cards were enough for a previously unknown team to overtake far more resource-hungry rivals and secure a spot near the top of the WorldArena table.

How a world model differs from a video generator

A robot world model isn't a generator of impressive-looking video — it's an engine for predicting physics. What's required of it isn't a "crisp frame" but a correct answer: where an object will end up after being pushed, whether the gripper will hold onto it, what happens to an object hidden behind an obstacle. This distinction is fundamental for robotics — an action can only be planned when the model correctly predicts its consequences. According to the description from 机器之心 (Jiqizhixin), it's precisely on scenes involving contact and occlusion that current systems fail, and it's precisely there that Moke's model showed an advantage on the WorldArena benchmark, having been trained on just 32 Zhenwu810E accelerators.

What this means

The rise of the "dark horse" Moke to the top of WorldArena signals a shift in the rules: breakthroughs in robotic world models are increasingly determined by architecture and data rather than cluster size. And success with just 32 Zhenwu810E accelerators shows that Chinese AI hardware is already capable of handling tasks that until recently were considered the domain of massive imported GPU clusters.