World models have finally entered the real-time era. While the industry is still struggling with 5FPS or 10FPS, a Chinese team has directly pushed real-time interactive world models to 50FPS. More importantly, they did not use NVIDIA GPUs.
Moxin Technology, a company focused on 4D world model R&D and industrialization, in collaboration with the team of Academician Pan Yunhe at Zhejiang University, has released MoWorld — the world's first Flash World Model, also the first real-time interactive world model built fully on domestic NPUs across the entire stack.
From training, distillation to deployment, the entire pipeline runs through a closed loop of domestic computing power; inference costs are 70% lower than comparable GPU-based solutions.
The inflection point for world model industrialization may arrive earlier than anyone anticipated.
The Real-Time Problem Stalling the Whole Industry
Punctured by a Chinese Team
If you've tried today's mainstream world models, you probably share a common feeling:
- They can be watched, but not played.
- Robots require real-time decision-making.
- Games require real-time feedback.
- Digital worlds require real-time simulation.
Past studies have shown that with a frame rate below 30FPS, all sense of immersion is shattered.
This is why for a long time, world models have remained in the lab, struggling to truly enter industrial applications.
Real-time performance was once the final insurmountable chasm on the path to commercializing world models.
Now, this chasm has been decisively crossed by a Chinese team.
50FPS!
World Models Have Truly Entered the Real-Time Era for the First Time
The technical report has been released. Weights and code will be open-sourced soon, and services will be provided to the public based on domestic NPU super-nodes.
Technical Report: https://moxin-tech.github.io/moworld/
Moxin Technology has introduced the concept of Flash World Model for the first time. MoWorld is the industry's first "Flash World Model" to achieve >50FPS inference, and the first low-cost real-time interactive world model built fully on domestic NPUs across the entire stack.
MoWorld has, for the first time, achieved a complete closed loop on a full-stack domestic computing power platform — from training and distillation to real-time inference deployment. Meanwhile, under typical inference configurations, costs are 70% lower than comparable GPU-based solutions.
While Everyone Else is Stockpiling GPUs
They Chose a Different, Harder but Right Path
From the beginning, MoWorld was not built around GPUs. Instead, they chose a more difficult and less-tried route — a full-stack domestic NPU approach.
This means it's not simply migrating the model to domestic chips for inference. Every stage, from data and training to distillation and inference deployment, has been redesigned around domestic NPUs.
First, data.
Compared to video generation models, world models require not just videos and text, but also information like camera trajectories. Internet videos are far from sufficient for training needs.
To address this, MoWorld established a complete data production and governance system, upgrading it into a data engine serving world models.
Leveraging Moxin's years of technical expertise in 3D/4D modeling, the data pipeline behind MoWorld is fully self-collected and fully 3D annotated, labeling not only cameras but also object geometric dimensions and spatial structures. This information forms the solid data foundation for MoWorld.
More challenging difficulties arose during training and inference.
Addressing the hardware characteristics of domestic NPUs, MoWorld redesigned its training system, introducing techniques like ultra-dense attention parallelism and long-sequence token parallelism. This greatly alleviated the memory pressure caused by ultra-long video training, enabling world models to possess 2000-frame long-term training and inference capability for the first time.
During the inference stage, the team continued system-level optimizations for domestic NPUs, including pipeline execution, hierarchical sequence parallelism, and dynamic mixed-precision quantization.
Ultimately, a 14B parameter MoE world model achieved an extreme real-time inference speed exceeding 50FPS on the Huawei Ascend 910C CloudMatrix384 NPU platform.
More importantly, under typical inference configurations, inference costs are 70% lower than comparable GPU-based solutions.
For the entire industry, this is not just a new engineering path, but a new industrialization strategy.
The Priciest Door for World Models Has Been Pushed Open
For a long time, world model development has been plagued by a contradiction.
Models are getting more powerful.
Costs are also getting higher.
Previously, deploying a world model meant high GPU investments, complex cluster maintenance, and hard-to-replicate deployment costs.
Now, the same world model capabilities can run on domestic computing power platforms with greater cost advantages.
This not only changes how models run, but also alters the path for world models moving towards industrialization.
For enterprises, this means lower deployment barriers, faster application validation, and easier large-scale replication.
For the entire industry, it means world models are beginning to shift from "can be made" to "can be used," and further to "affordable to use."
What truly drives a technology to change an industry is never the highest record in a lab.
But the first time it enables more people to actually use it.
World Models Step Out of the Lab
Which Industries Will Face Transformation
For the past few years, world models have been considered a future technology.
But the future must eventually return to reality.
When real-time interaction becomes possible and deployment costs begin to drop, the true value of world models is just starting to be unleashed.
The first changes will occur in industries most reliant on real-world understanding and real-time feedback.
Gaming & Interactive Entertainment: Real-Time Interaction, Free Exploration
MoWorld supports complete 6-DoF (Degrees of Freedom) camera control. Users can achieve cinematic and game-level immersive exploration experiences using just W/A/S/D keys and mouse.
Scenes are realistic, high-definition, supporting 1080P and higher resolutions. Comprehensive support is provided for natural landscapes, anime, and game animations.
Embodied AI & Autonomous Driving: Virtual Training, Real-World Validation
World models have become a key bridge connecting generative AI with embodied intelligence.
MoWorld can provide low-cost, high-fidelity "digital training grounds" for robots and autonomous driving systems. It is the most promising world simulator in the industry, combining simulation value and economic value, capable of providing vast amounts of high-precision environments for all autonomous driving teams. This allows AI to learn how to interact with real physical environments within virtual worlds.
Film & TV Production: Director's Cinematography, Real-Time Previsualization
Traditional storyboarding for film and TV requires long rendering cycles.
MoWorld allows creators to freely adjust perspectives, preview visual effects in real-time, and precisely edit shot compositions within generated virtual worlds. Camera control is smooth, supporting director-level cinematography beyond imagination.
Digital Twins & 3D Reconstruction: Spatial Reconstruction, Accurate Restoration
Videos generated by MoWorld possess surpassing industry standards in geometric consistency, enabling direct use for 3D reconstruction of indoor scenes — high precision, stable structure, and excellent spatial consistency are the distinguishing effects setting MoWorld apart from peers.
This provides solutions combining high precision and cost-effectiveness for scenarios like digital twins, architectural visualization, virtual showrooms, and immersive games.
The World Model Represented by MoWorld
Is Now Standing at the DeepSeek Moment for Physical AI
In recent years, AI has completed successive leaps from text generation, image generation to video generation. Each technological leap has spawned a new generation of industry leaders.
As AI now moves towards real-time interactive world models, a new window is opening.
Compared to the already competitive landscape of large language models and video generation models, world models remain in the early industrial stage, with global exploration of engineering implementation paths and industry standards far from established.
This means domestic world models have encountered a rare "equal starting line" opportunity — not only the chance to compete but also to participate in defining the technical standards for the next generation of spatial intelligence.
The significance of MoWorld lies not merely in achieving over 50FPS real-time interaction, nor solely in reducing inference costs by 70% compared to similar-scale GPU solutions.
More importantly, it validates a question the industry has long explored but never proven:
A full-stack domestic computing power solution can indeed support world models moving towards real-time interaction and industrial deployment.
This means the competition for world models is shifting from "who has the larger model" to "who can truly enter the real world."
Moxin Technology, behind MoWorld, with its unique model capabilities and outstanding industrialization progress, recently secured over one billion USD in financing from national strategic reserve capital, well-known Middle Eastern dollar institutions, leading market-oriented funds, and over ten industrial capital firms. Prior to this, Moxin Technology had received investment from Huawei's Hubble Investment and Lenovo Holdings' funds.
The era truly belonging to world models is beginning with Flash World Model and will rapidly advance from here.
This article is from the WeChat public account "Xin Zhi Yuan," author: ASI Revelation
