Just now, the world’s first ultra-high frame world model was born, with Nvidia content of 0 and skyrocketing to 50 frames per second

MoWorld, a team led by Academician Pan Yunhe from Zhejiang University, has released MoWorld – the world’s first Flash World Model and the first real-time interactive world model built on a domestic NPU, which focuses on the development and industrialization of 4D world models.

From training, distillation to deployment, the entire process runs through the closed loop of domestic computing power; The inference cost is reduced by 70% compared to GPU solutions of the same scale.

The industrialization turning point of the world model may come earlier than everyone expected.

Stuck the real-time challenges of the entire industry
Exposed by the Chinese team
If you have experienced today’s mainstream world models, you are likely to have a common feeling:

Can see, but cannot play.
Robots require real-time decision-making.
The game requires real-time feedback.
The digital world requires real-time inference.
Previous studies have shown that if the frame rate is below 30FPS, all immersion will be shattered.

This is also why for a long time in the past, the world model remained in the laboratory and was difficult to truly enter industrial scenarios.

Real time performance was once the last hurdle on the road to commercializing world models.

Now, this natural barrier has been conquered by a Chinese team in one fell swoop.

50FPS！
The world model has truly entered the real-time era for the first time
The technical report has been released and will soon open source the weights and code, and provide services to the public based on domestic NPU supernodes.

Technical report: https://moxin-tech.github.io/moworld/

MoXin Technology first proposed the concept of Flash World Model. MoWorld is the industry’s first “Flash World Model” to achieve>50FPS inference, and also the first low-cost real-time interactive world model built on domestic NPU with a full stack.

MoWorld has achieved a complete closed-loop from training, distillation to real-time inference deployment on a full stack domestic computing platform for the first time, while reducing costs by 70% compared to GPU solutions of the same scale under typical inference configurations.

When everyone is stacking GPUs
They chose another difficult but correct path
From the beginning, MoWorld did not build around GPUs, but chose a more difficult and less explored path – full stack domestic NPU.

This means that it is not simply transferring the model to domestic chips for inference, but rather redesigning every step from data, training, distillation to inference deployment around domestic NPUs.

Firstly, it is data.

Compared to video generation models, world models require not only video and text, but also information such as camera trajectories. Internet video is far from meeting training needs.

To this end, MoWorld has established a complete data production and governance system and upgraded it to a data engine that serves the world model.

Based on the years of 3D/4D modeling technology accumulation of the MoWorld team, the data pipeline behind MoWorld is a fully self collected and fully 3D annotated pipeline, which not only annotates cameras but also annotates the geometric dimensions and spatial structure of objects. These information build a solid data foundation for MoWorld.

The more difficult challenges occur during the training and reasoning stages.

In response to the hardware characteristics of domestically produced NPUs, MoWorld has redesigned its training system, introducing technologies such as ultra dense attention parallelism and long sequence token parallelism to significantly alleviate the memory pressure caused by ultra long video training, enabling the world model to have the ability to train and reason for 2000 frames for the first time.

In the inference stage, the team continued to perform a series of system level optimizations on domestic NPUs, including pipeline execution, hierarchical sequence parallelism, and dynamic mixed precision quantization.

Finally, a 14B parameter MoE world model achieved a real-time inference speed of over 50FPS on the Huawei Ascend 910C CloudMatrix384 NPU platform.

More importantly, in a typical inference configuration, the inference cost is reduced by 70% compared to GPU solutions of the same scale.

For the entire industry, this is not only a new engineering route, but also a new industrialization idea.

The most expensive door of the world model has been pushed open
For a long time, there has been a contradiction in the development of world models.

The model is becoming stronger and stronger.

The cost is also increasing.

In the past, deploying a world model meant high GPU investment, complex cluster maintenance, and difficult to replicate deployment costs.

Now, the same world model capability can run on more cost-effective domestic computing platforms.

This not only changes the way the model operates, but also changes the path of the world’s models towards industrialization.

For enterprises, this means lower deployment thresholds, faster application validation, and easier scalability for replication.

For the entire industry, this means that the world model is starting to move from ‘capable’ to ‘usable’, and then to ‘affordable’.

What truly drives a technology to change an industry is never the highest record in the laboratory.

But it’s the first time that more people can truly use it.

World models step out of the laboratory
Which industries will undergo changes
In the past few years, the world model has been considered as the future technology.

But in the future, we will eventually return to reality.

When real-time interaction becomes possible, when deployment costs begin to decrease, the true value of the world model is just beginning to be unleashed.

The industries that rely most on real-world understanding and real-time feedback will be the first to undergo changes.

Games and interactive entertainment: real-time interaction, free exploration

MoWorld supports full 6-degree-of-freedom camera control, allowing users to experience immersive roaming at both film and gaming levels through W/A/S/D and mouse.

Realistic and high-definition scenes, supporting resolutions of 1080P and above. Whether it is natural scenery, anime or game animation, it is fully supported.

Embodied intelligence and autonomous driving: virtual training, real verification

The world model has become a key bridge connecting generative AI and embodied intelligence.

MoWorld can provide low-cost, high fidelity “digital drill ground” for robots and auto drive system. It is the most potential world simulator with both simulation value and economic value in the industry. It can provide a large number of high-precision environments for all intelligent driving teams, so that AI can learn how to interact with the real physical environment in the virtual world.

Film and television creation: director’s cinematography, real-time rehearsal

Traditional film and television storyboards require a long rendering cycle.

MoWorld allows creators to freely adjust their perspective, preview image effects in real-time, and edit lens images accurately in the generated virtual world. The lens control is silky smooth and supports director level camera movements that go beyond imagination.

Digital twin and 3D reconstruction: spatial reconstruction, precise restoration

The videos generated by MoWorld have geometric consistency beyond the industry and can be directly used for 3D reconstruction of indoor scenes – high accuracy, stable structure, and good spatial consistency are the significant effects that distinguish MoWorld from its peers.

This provides a high-precision and cost-effective solution for scenarios such as digital twins, architectural visualization, virtual exhibition halls, and immersive games.

The world model represented by MoWorld
At this moment, we are standing at the DeepSeek moment of physical AI
In the past few years, AI has made continuous leaps from text generation, image generation to video generation, and each technological leap has spawned a new group of industry leaders.

And as AI begins to move towards real-time interactive world models, a new window is opening.

Compared to the large language models and video generation models that have gradually formed a competitive landscape, the world model is still in the early stages of the industry, and the world is exploring the path of engineering implementation, with industry standards far from being formed.

This means that domestic world models have a rare opportunity to “run on the same track” – not only to compete, but also to participate in defining the technical standards for the next generation of space intelligence.

The significance of MoWorld lies not only in achieving real-time interaction at over 50FPS, but also in reducing inference costs by 70% compared to GPU solutions of the same scale.

More importantly, it proves a problem that the industry has been exploring but has yet to verify:

Full stack domestic computing power can also support the world model towards real-time interaction and industrial deployment.

This means that the competition for world models is shifting from ‘who has the bigger model’ to ‘who can truly enter the real world’.

MoWorld’s MoXin Technology, with its unique modeling capabilities and outstanding industrialization progress, has recently obtained financing of over 100 million US dollars from national strategic reserve capital, well-known US dollar institutions in the Middle East, top market-oriented funds, and more than ten industrial capital companies; Prior to this, Magic Core Technology had obtained investments from Huawei’s Hubble Investment and Lenovo Holdings’ fund.

The era that truly belongs to the world model, starting from the Flash World Model, will move forward rapidly.