MoWorld - The World's First High-Frame-Rate Interactive World Model Launched by MoXin Technology

What is MoWorld

MoWorld is the world's first high-frame-rate interactive world model launched by MoXin Technology. Built on fully domestic Huawei Ascend NPU full-stack optimization, it achieves up to 50 FPS real-time inference with 70% lower cost than GPU. The model supports precise camera trajectory control and 4D spatial perception, and can generate long-duration, spatially consistent large-scale scene videos such as urban, indoor, and driving scenes. MoWorld is suitable for film and television previsualization, real-time cloud gaming, embodied simulation, autonomous driving, and 3D scene reconstruction.

Main Features of MoWorld

• High-Frame-Rate Real-Time Interaction: 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 Modeling: The model integrates precise camera trajectories and spatial geometric information to achieve 4D dynamic world simulation with spatial perception.
• Long-Duration Spatial Consistency: In large-scale bird's-eye-view scenes such as cities, highways, and deserts, it maintains the stability of building layouts, road directions, and distant details, with no drift in cross-view geometric relationships.
• Joint Camera-Storyline Control: Supports coordinated control of director-level camera movement and narrative events. The generated videos can be directly used for 3DGS scene reconstruction and film and television previsualization workflows.
• Low-Cost Domestic Deployment: Based on full-stack native optimization for Huawei Ascend NPU, inference cost is 70% lower than GPU, proving for the first time that domestic computing power can support real-time deployment of large-scale world models.
• Multi-Scene Generalized Generation: Training data covers both the real world and the virtual world, balancing richness and cross-domain generalization capability.

Technical Principles of MoWorld

• Data Engine: Covers both the real world and the virtual world, building high-quality training assets through multi-dimensional quality filtering such as 3D geometric consistency, trajectory accuracy, and multi-view stability.
• Pre-Training: Ultra-dense attention parallelism + input token parallelism, supporting foundation model training up to 2 minutes long.
• Distillation Optimization: Compressed into an autoregressive student model with 4-step real-time inference, introducing a Memory mechanism with global anchors and camera stability consistency, skipping teacher trajectory sampling initialization, and significantly reducing distillation cost.
• Real-Time Inference: Three-layer collaborative optimization—pipeline layer (condition encoding reuse, on-demand module loading), parallelism layer (multi-NPU ring communication, hierarchical sequence parallelism), and operator layer (mixed-precision dynamic quantization, HBM-efficient attention frontend).

How to Use MoWorld

MoWorld is in the early release stage, and the weights, code, and online service will be opened soon.

Core Advantages of MoWorld

• The World's First Flash World Model: Pioneers the concept of an ultra-high-frame-rate interactive world model, achieving up to 50 FPS real-time inference and bringing world models from non-real-time generation into the era of real-time interaction.
• Fully Domestic NPU Full-Stack Native: For the first time, it opens up the complete technical chain from pre-training and autoregressive distillation to real-time inference on Huawei Ascend NPU, proving that domestic computing power can independently support large-scale world model training and deployment.
• 70% Lower Inference Cost: After extreme NPU optimization, deployment cost is significantly reduced compared with GPU, making high-frame-rate world models economically feasible for large-scale implementation for the first time.
• Low Cost Without Sacrificing Results: The 14B-parameter MoE architecture achieves industry-TOP-level image quality, spatial consistency, and temporal stability, and can still maintain precise structures of buildings, roads, and distant views in large-scale bird's-eye-view scenes.
• Long-horizon spatial consistency: By introducing global anchors and a camera-stable Memory mechanism, even after large-range camera movements such as high-altitude dives and rapid pull-ups, cross-view geometric relationships remain stable without drift.
• Joint camera-storyline control: Supports coordination between precise camera trajectories and narrative events. Generated videos can directly reconstruct 3DGS scenes, closing the application loop for film and television previsualization, embodied simulation, and digital twins.

MoWorld project address

• Project website：https://moxin-tech.github.io/moworld/
• Technical paper：https://moxin-tech.github.io/moworld/MoWorld.pdf

Comparison of MoWorld with similar competing products

MoWorld application scenarios

• Film and television production and shot previsualization: Creators can focus on character actions and camera language, significantly reducing background drift, viewpoint jumps, and scene identity inconsistency issues.
• Real-time cloud gaming and interactive content: Users explore generated virtual worlds in real time through continuous camera control, supporting low-latency roaming, camera control, and immersive interactive experiences.
• Indoor scene reconstruction and digital twins: Use videos generated by MoWorld for 3D reconstruction of indoor scenes, with high target detail, stable reconstruction, and good spatial consistency, and can directly output digital scene assets.
• Embodied intelligence simulation: The high-frame-rate Flash World Model enables robotic systems to perform environmental perception and decision-making at unprecedented speed, providing a real-time physical interaction simulation environment for embodied intelligence.
• Autonomous driving training and validation: By generating diverse driving scenarios in real time (urban, highway, desert, etc.), it supports rapid iteration and safety validation of autonomous driving algorithms in dynamic, large-scale environments.