Moke Storms WorldArena's Top Ranks With a Robot World Model Trained on 32 Zhenwu 810E Chips

China's Moke has unveiled a world model for robots that reached the top of the WorldArena benchmark after training on just 32 Zhenwu 810E accelerators, according to a report published on July 30, 2026 by 机器之心 (Jiqizhixin).

What problem is Moke solving?

Today's robot world models struggle to hold onto real-world physics: a gripper fails to lift an object, a robotic arm does not interact with its environment, and once objects overlap in the frame, the prediction falls apart completely. The problem is not "pretty pictures." AI-generated video of the future looks sharp frame by frame, but diverges from reality at the moment of contact. That is the failure Moke's model targets: it aims for physically credible predictions of what happens when a robot interacts with objects. According to 机器之心 (Jiqizhixin), current world models repeatedly fail on exactly these scenes: grasping, contact and occlusion.

"On screen, the simulated robotic arm follows a perfect trajectory,

but in the real world the glass is either still standing on the table, or has already been knocked over by the gripper," reads a passage from the 机器之心 (Jiqizhixin) article.

Key facts from the report:

• The developer is the Chinese team Moke (莫刻), a previously little-known dark horse
• The model is a world model for robots that rose to the top of the WorldArena ranking
• Hardware: just 32 Zhenwu 810E (真武810E) accelerators
• Publication date: July 30, 2026; source: 机器之心 (Jiqizhixin)
• The task: accurately predicting grasping, interaction with the environment, and scenes with overlapping objects

Why do just 32 chips matter?

The main intrigue is compute efficiency. Moke reached the top of WorldArena using only 32 Zhenwu 810E accelerators, whereas climbing to the top of comparable 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. Zhenwu 810E is a Chinese AI accelerator, and betting on it underscores the push for technological independence from imported GPUs. The figure "32" is highlighted in the article's headline as a record: that is how many cards a previously unknown team needed to outperform more resource-hungry rivals and secure a place in the upper part of the WorldArena table.

How does a world model differ from a video generator?

A robot world model is not a generator of spectacular video but a physics-prediction engine. What it needs to deliver is not a "sharp frame" but a correct answer: where an object ends up after being pushed, whether the gripper will hold it, what happens to an object hidden behind an obstacle. The difference is fundamental for robotics: an action can only be planned if the model correctly predicts its consequences. According to 机器之心 (Jiqizhixin), current systems fail precisely on contact and overlap scenes, and this is where Moke's model showed an advantage on the WorldArena benchmark, trained on just 32 Zhenwu 810E accelerators.

What does this mean?

The appearance of the dark horse Moke at the top of WorldArena points to a change in the rules: breakthroughs in robot world models are increasingly determined by architecture and data rather than cluster size. And the success on 32 Zhenwu 810E accelerators shows that Chinese AI hardware can already handle tasks that were considered the domain of giant imported GPU clusters until recently.