Moxin Technology launches MoWorld 2.0, a 4D world model trained and deployed natively on Huawei Ascend

Moxin Technology has released MoWorld 2.0, a 4D world model trained and deployed natively on Huawei Ascend hardware, adding physical simulation to the scenes it generates. The company unveiled it at a World Internet Conference forum in Wuzhen on November 4, IT Home reported. It follows MoWorld 1.0, released in July.

What changed in 2.0

According to Moxin, the main step up is physics. The model folds gravity, elasticity, buoyancy, lighting and heat, along with the properties of rigid bodies, soft bodies, fluids and deformable materials, into how it builds and interacts with spaces. Its output is 4D Gaussian assets that can be used for simulation, rendered in real time and reused repeatedly.

Moxin says the model generates not only the appearance of a space and its objects but also reasons in real time about how objects interact, how external forces act on them and how the environment changes. In one demonstration, a breeze turns the pages of a book and a bookmark falls to the floor. That scene requires accounting for wind strength and direction, paper flexibility, the bookmark's weight and the rigidity of the floor. Other demos show scenes that can be edited and interacted with, and time that can be sped up or slowed down.

Ascend-native training and inference

MoWorld 2.0 was trained and deployed on a fully domestic Ascend computing base. Moxin says it worked with Huawei teams on long-sequence training, multi-NPU scheduling and inference system optimization. By its own tests and estimates, the company says inference cost can fall to as little as 30% of a GPU setup of the same scale. This is a company claim, not an independent measurement.

Speed figures (company-reported)

• Video inference: peak of about 50 frames per second in version 1.0, about 70 frames per second in 2.0.
• Rendering of 3D and 4D Gaussian assets: above 500 frames per second.
• Phone tests: about 100 frames per second.

Moxin previously brought a lightweight MoWorld model to Huawei's Mate 90 phone.

Target applications

Moxin says it is working with partners on validation and deployment in several areas:

• Embodied AI: virtual environments with spatial structure and basic physical feedback, so robots can train and retry in simulation.
• Autonomous driving: generation of long-tail scenes such as rain, snow, potholes and sudden obstacles.
• Industrial digital twins: factory lines and mines.
• Urban planning and virtual tourism.
• Film and game production: reusable scene assets that support free camera movement, object editing and weather changes.

Context window, benchmark scores and parameter count were not part of the announcement as reported. Pricing not yet disclosed.

What this means

MoWorld 2.0 shifts Moxin's model from generating how a space looks toward modeling how objects in it behave under forces. Because the output is 4D Gaussian assets that render in real time and can be reused, the same scene can serve simulation and production work.

The Ascend-native claim is the second point to watch. Moxin says the model was trained and deployed entirely on domestic Ascend hardware, with a reported inference cost as low as 30% of a same-scale GPU setup. That figure comes from Moxin's own tests and estimates, so it should be treated as unverified until partners or third parties publish comparable measurements. The same applies to the application areas: Moxin describes validation with partners as ongoing, not completed.