MoWorld - The world's first high frame rate interactive world model launched by Mochip Technology

What is MoWorld?

MoWorld, launched by Mochip Technology, is the world's first high-frame-rate interactive world model. Based on a fully optimized, domestically developed Huawei Ascend NPU stack, it achieves up to 50 FPS real-time inference at a cost 70% lower than GPUs. The model supports precise camera trajectory control and 4D spatial awareness, generating long-term, spatially consistent large-scale scene videos, including urban, indoor, and driving scenarios. MoWorld is suitable for film and television pre-visualization, real-time cloud gaming, embodied simulation, autonomous driving, and 3D scene reconstruction.

MoWorld's main functions

• High frame rate real-time interactionIt supports up to 50 FPS real-time inference, bringing world models from non-real-time generation into the era of real-time interaction.
• 4D World ModelingThe model integrates precise camera trajectories and spatial geometric information to achieve a 4D dynamic world simulation with spatial awareness.
• Long-term spatial consistencyMaintain the stability of building layout, road orientation, and distant details in large-scale bird's-eye view scenes such as cities, highways, and deserts, and ensure that cross-view geometric relationships do not drift.
• Camera-Story Joint ControlIt supports director-level coordinated control of camera movement and narrative events, and the generated videos can be directly used for 3DGS scene reconstruction and film and television pre-visualization processes.
• Low-cost domestic deploymentBased on the full-stack native optimization of Huawei Ascend NPU, the inference cost is reduced by 70% compared to GPU, proving for the first time that domestic computing power can support the real-time deployment of large-scale world models.
• Multi-scenario generalization generationThe training data covers both the real and virtual worlds, balancing richness and cross-domain generalization ability.

MoWorld's technical principles

• Data EngineIt covers both the real and virtual worlds, and constructs high-quality training assets through multi-dimensional quality screening, including 3D geometric consistency, trajectory accuracy, and multi-view stability.
• Pre-training: Ultra-dense attention parallelism + input token parallelism, supporting base model training of up to 2 minutes.
• Distillation optimizationThe autoregressive student model is compressed to 4 steps of real-time inference, and a memory mechanism that ensures global anchor point and camera stability is introduced. It skips teacher trajectory sampling initialization and significantly reduces distillation costs.
• Real-time inferenceThree-layer collaborative optimization: pipeline layer (conditional coding reuse, on-demand module loading), parallel layer (multi-NPU ring communication, hierarchical sequence parallelism), and operator layer (mixed precision dynamic quantization, HBM high-efficiency attention front end).

How to use MoWorld

MoWorld is in its early stages of release; weights, code, and online services will be available soon.

MoWorld's core advantages

• The world's first Flash World ModelIt pioneered the concept of an ultra-high frame rate interactive world model, achieving up to 50 FPS real-time inference, bringing world models from non-real-time generation into the era of real-time interaction.
• Purely domestically produced NPU full-stack nativeThis achievement marks the first time that a complete technical chain from pre-training, autoregressive distillation to real-time inference has been established on Huawei's Ascend NPU, proving that domestic computing power can independently support the training and deployment of large-scale world models.
• Reasoning costs reduced by 70%After extreme optimization with NPU, the deployment cost is significantly reduced compared to GPU, making the high frame rate world model economically feasible for large-scale deployment for the first time.
• Low cost without sacrificing effectivenessThe 14B parameter MoE architecture achieves top-tier performance in the industry in terms of image quality, spatial consistency, and temporal stability, maintaining the precise structure of buildings, roads, and distant views even in large-scale bird's-eye view scenes.
• Long-term spatial consistencyBy introducing a global anchor point and camera stabilization memory mechanism, the cross-view geometric relationship remains stable and does not drift even during large-scale camera movements such as high-altitude dives and rapid ascents.
• Camera-Story Joint ControlIt supports precise camera trajectory and narrative event coordination, and the generated video can directly reconstruct 3DGS scenes, opening up a closed loop of applications for film and television pre-visualization, embodied simulation and digital twins.

MoWorld's project address

• Project official websitehttps://moxin-tech.github.io/moworld/
• Technical Papers: https://moxin-tech.github.io/moworld/MoWorld.pdf

MoWorld's Competitive Product Comparison

MoWorld Application Scenarios

• Film production and pre-visualizationCreators can focus on character movements and camera language, significantly reducing issues such as background drift, perspective jumps, and inconsistencies in scene identity.
• Real-time cloud gaming and interactive contentUsers can explore the generated virtual world in real time through continuous camera control, supporting low-latency roaming, lens control, and immersive interactive experiences.
• Indoor scene reconstruction and digital twinMoWorld-generated videos are used for 3D reconstruction of indoor scenes, resulting in high target precision, stable reconstruction, and good spatial consistency. Digital scene assets can be directly output.
• Embossed Intelligent SimulationHigh frame rate Flash World Model enables robotic systems to perceive and make decisions about their environment at unprecedented speeds, providing a real-time physical interaction simulation environment for embodied intelligence.
• Autonomous driving training and verificationBy generating diverse driving scenarios in real time (city, highway, desert, etc.), it supports the rapid iteration and safety verification of autonomous driving algorithms in dynamic, large-scale environments.