MoWorld 2.0 Debuts at Wuzhen: a 4D World Model Trained Entirely on Huawei's Ascend

Chinese startup MoXin Technology used the World Internet Conference in Wuzhen on November 4 to launch MoWorld 2.0, the second version of its 4D world model, trained and deployed natively on Huawei's Ascend compute stack. Founder and chief executive Chen Tianren presented the model alongside representatives of three national AI application pilot bases covering embodied intelligence, culture and tourism, and spatiotemporal geoinformation. The company is a participant in Huawei's Kunpeng-Ascend seed program for early-stage compute startups.

The upgrade pushes past the first version, released in July, in one specific direction: physics. MoWorld 2.0 folds gravity, elasticity, buoyancy, lighting and heat — plus rigid, soft, fluid and deformable body models — into spatial understanding and interaction. The company's illustrative case is a bookmark falling from a book page in a simulated breeze: getting that right means reasoning jointly about wind force and direction, paper flexibility, the bookmark's weight, floor rigidity and contact relationships. Output is a set of 4D Gaussian splat assets that can be simulated, rendered in real time, stored and reused rather than re-generated.

Speed claims are substantial but company-reported and not independently verified. Video inference rises from about 50 frames per second in version 1.0 to about 70. Rendering of 3D and 4D Gaussian assets runs above 500 frames per second, and on-phone rendering reaches roughly 100 frames per second in the company's tests. MoXin also says its Ascend-optimized inference path can bring inference cost down to as little as 30 percent of an equivalent GPU-based setup.

The Ascent-native framing is the story's other half. The model is trained and served on Ascend hardware, with joint optimization work on long-sequence training, multi-NPU scheduling and inference systems. In October the company's consumer app — which turns a few pet photos into a 4D digital pet avatar users can rotate and interact with — appeared at a Huawei flagship phone launch, an early sign the pipeline runs from Ascend datacenter to handset. Consumers, robot developers, carmakers and studios would all draw on the same model.

The company lists five application fronts in various stages of delivery: embodied AI training, where robots rehearse in physically plausible virtual environments before touching real hardware; autonomous driving, where the model generates long-tail scenarios such as rain, potholes and sudden obstacles on demand; industrial digital twins for factory line re-planning; film and game production with reusable scene assets; and consumer tourism and immersive content. Commercial deployment remains early, and the pilot-base partnerships announced in Wuzhen are validation programs rather than revenue commitments.

MoXin says it has closed five funding rounds this year totaling 2 billion yuan, with investors spanning state-backed funds, industrial capital and technology investors across chips, compute, devices and content platforms. That pace is notable for a company founded on a research bet that world models — not larger video generators — are the durable asset in spatial AI.

The strategic point cuts across the field. Most headline world-model efforts, from Runway's Praxis-1 to Odyssey and Decart, run on Nvidia silicon; MoWorld is being positioned as proof that a full spatial-AI stack — model, physics and inference — can be rebuilt on a domestic compute base. Whether the physics claims survive outside scrutiny will matter to researchers first. But the cost claim, if it holds, speaks to the buyers: whoever can serve a world model at a third of the going rate gets to decide what "interactive simulation" costs as a component.