Maxchip releases MoWorld 2.0 4D world model on November 4 natively on Ascend, cutting inference cost to 30 percent
MoWorld 2.0 folds gravity, elasticity, buoyancy, light, heat and rigid, soft and fluid properties into spatial modelling, outputting simulatable, real-time renderable 4D Gaussian assets.
On November 4, at the World Internet Conference forum on new quality productive forces in AI, Maxchip formally released MoWorld 2.0, a 4D world model running natively on Huawei Ascend. The company said that following MoWorld 1.0 in July, version 2.0 lifts model capability by folding gravity, elasticity, buoyancy, light, heat and properties of rigid, soft, fluid and deformable bodies into spatial modelling and interaction capability, outputting 4D Gaussian assets that can be simulated, rendered in real time and reused repeatedly.
Through modelling multiple physical elements, MoWorld 2.0 does more than generate space and object appearance, performing real-time reasoning about interactions between objects, external forces and environmental change. Simulating a breeze turning book pages and a bookmark falling to the floor, for example, requires simultaneously accounting for wind strength and direction, paper flexibility, bookmark weight, and floor rigidity and contact relationships.
MoWorld 2.0 was trained and deployed on a fully domestic Ascend compute base, with Maxchip and Huawei teams continuing joint work on long-sequence training, multi-NPU scheduling and inference system optimization. According to the company's own testing, inference cost for the approach can fall as low as 30 percent of an equivalent GPU solution. Folding physical properties into controllable generation is the key technical line separating embodied world models from video generation models.