MoCore launches 4D world model MoWorld 2.0, cutting inference cost to 30% of GPU plans

At the World Internet Conference forum, Chinese tech firm MoCore launched MoWorld 2.0, a 4D world model built on Huawei's Ascend chips. It brings physics like gravity, light and fluids, plus object traits like rigid and soft bodies, into spatial modeling, creating 4D assets that support simulation, real-time rendering and repeated use.

Compared with version 1.0 released in July, the new model can also reason in real time about how objects interact and how environments change. For example, simulating wind flipping a book page means calculating wind force, paper flexibility and contact with the floor.

The model was trained and deployed on fully domestic Ascend computing power, and its inference cost can drop to as low as 30% of similar GPU solutions. Video inference runs at about 70 fps, 3D/4D asset rendering tops 500 fps, and mobile rendering hits about 100 fps.

The technology is being applied to robot training, autonomous driving, industrial digital twins, and film and game production, letting robots practice and retry in virtual spaces while simulating complex road conditions.