MoWorld

The world's first high frame rate interactive world model released by Moxin Technology, a purely domestic NPU achieves up to 50FPS real-time world simulation

video generation

https://moworld.ai/

Visit Website

In-depth Report

MoWorld is the "Flash World Model" officially released by Moxin Technology on July 7, 2026. It is officially defined as the world's first high frame rate interactive world model. Its core breakthrough is not "the ability to generate a video", but the ability to generate real-time interactive video worlds at up to 50FPS on a purely domestic Huawei Ascend NPU, with inference costs only 30%–50% of that of GPU solutions of the same scale. This has been called the "DeepSeek moment" in the world model field by many media.
Moxin Technology is a company focusing on the research, development and industrialization of 4D world models. It was founded in 2021 by Chen Tianrun, a direct PhD student at the School of Computer Science at Zhejiang University (a "post-00s" entrepreneur), and studied under Pan Yunhe, an academician of the Chinese Academy of Engineering and former president of Zhejiang University. On the same day as the release of MoWorld, the company announced the completion of nearly 1 billion yuan in Series A financing. Investors include national strategic capital, Middle East dollar institutions and more than a dozen industrial capitals. It has previously received strategic shares from Huawei Hubble Investment, Legend Holdings' funds, and Richhan Micro. The team’s goal is to reach the first echelon of the global universal world model prototype in the next 3–5 years.
MoWorld accepts three types of conditions: "first frame + text + camera trajectory", generates a future world state that conforms to the scene state and control input, and allows users to roam and explore the generated world in real time through a continuous control method similar to W/A/S/D plus mouse. It supports complete 6 degrees of freedom camera control and resolutions of 1080P and above, whether it is natural scenery, two-dimensional or game animation scenes. On the technical base, MoWorld uses the Wan2.2-A14B video basic model as its backbone, adopts the 14B parameter MoE (hybrid expert) architecture, and adds a new Camera Adapter to inject camera control. The training side built a 3D geometric consistency data engine covering the real world and the virtual world, and conducted cross-frame pre-training of the course (gradually extending from 125 frames to 2000 frames long); the inference side used autoregressive distillation to reduce the number of denoising steps from 50 steps to 4 steps, and then superimposed the three-level collaborative optimization of "pipeline layer + parallel layer + operator layer" to achieve up to 50FPS on the Huawei Ascend NPU super node. Real-time reasoning. "High frame rate" is the key to distinguishing it from ordinary video generation models - for real-time roaming, robot training, virtual shooting and game previews, low frame rate will directly destroy the sense of control. MoWorld advances the world model to the "accessible, controllable, low-latency feedback" range for the first time.
MoWorld is currently in the early stages of release, and weights, codes and online services "will be available soon." Official disclosures will provide services to the public based on domestic NPU super nodes. The technical report has been released, and follow-up plans are to open source the weights and code. Its business logic is not to sell models directly, but to use low-cost, high-fidelity world simulation capabilities as "spatial intelligence infrastructure" to export to industries such as film and television, games, embodied intelligence, autonomous driving, and digital twins.
Since the product has just been released and has not yet been fully open to public testing, there are still very few real end-user feedback samples. At present, it mainly comes from technical discussions between industry media and the developer community. One school of thought believes that the world model cannot open up the industry for a long time, and the root cause lies in the interactive frame rate and computing power bill - real-time deployment cannot be discussed when it is stuck in the screen generation stage, and MoWorld's 50FPS and "cost reduction of 70%" just respond to these two stuck points. The other group remains cautious, believing that it is now more like a technical route signal. The actual entry into the production process depends on the stability, resolution, control accuracy and material copyright boundaries of the code, weight and experience portal after the opening.
The industry generally compares it with real-time world models such as Tencent Hunyuan HYWorld 1.5, overseas Genie, and GameGen. In terms of frame rate, MoWorld's maximum 50FPS is significantly higher than most mainstream world models below 30FPS; in terms of cost, NPU full-stack native optimization reduces inference costs to 30%–50% of GPU solutions (some reports claim the reduction is as high as 70%). In terms of benchmark testing, MoWorld achieved the best Quality 82.73 and Average 85.22 in the VBench-I2V official subset, and the Average reached 90.31 in the internal camera/world model data set. The media commented that it proves that domestic NPU can independently complete the entire link of the world model from pre-training, distillation to real-time reasoning, and has the economic feasibility of industrial implementation.
The main disputes focus on three points. First, there is still a distance between the "domestic narrative" and the engineering reality - the weights and codes have not yet been open sourced, and the true usability needs to be verified. The second is the copyright and compliance boundaries commonly faced by world models: real-world videos are used to train and generate interactive scenes, and the commercial boundaries between material authorization and generated content have not yet been clarified. The third is the upper limit of real-time interaction experience: most of the existing demonstrations are fragment-level scenes. Whether the image quality attenuation, object drift and timing flicker under long-term continuous control can be completely solved still requires larger-scale public testing.
MoWorld is most suitable for teams with clear "controllable scene generation" needs: film and television storyboard and pre-show creators can use it to quickly lay out shot language and reduce background drift; game and digital twin teams can use it as a low-cost environment engine; embodied intelligence and autonomous driving teams can use it as a high-fidelity "digital training ground" to repeatedly train perception and decision-making in the virtual world. For ordinary users who only want to generate a single short video, the threshold is currently high. It is recommended to wait until online services and open source weights are opened before evaluating. If you are looking for similar but lighter solutions, you can pay attention to competing products such as Tencent HYWorld 1.5 and Google Genie series.
MoWorld advances the world model from "generable" to "interactive and deployable", using a purely domestic NPU to run through 50FPS real-time world simulation and significantly reducing costs. It is a benchmarking technological breakthrough in the field of spatial intelligence; whether it can truly enter the industry depends on the stability, ecology and service-oriented progress after open source.

User Reviews

TEmor

—

We do digital twins and have extremely high requirements for geometric consistency. MoWorld promotes that large-scale bird's-eye view scenes do not drift in geometry across viewing angles. This would be great if true. Directly producing 3D assets for indoor reconstruction can save a lot of manual modeling. But just watching the promotional video cannot convince anyone about the picture quality of the two-dimensional and game animation scenes.

EReed_66_986

—

After paying attention, there is basically no real end-user feedback at present. It is all discussed by industry media and developers. This is normal, after all, it has just been released and is not yet in public beta. It is recommended that teams who want to use it don’t get excited and wait until the VBench indicators in the technical report are independently reproduced by a third party and the online service is opened to run a round of real scenarios before deciding whether to access it.

orangefish486

—

As someone who works in the direction of autonomous driving simulation, what I care most about is its ability to generate diverse driving scenarios. Cities, highways, and deserts can all be generated in real time, which theoretically can provide our algorithms with a large number of high-precision virtual environments. But the hidden dangers are also obvious - if the distribution deviation of the generated content is not clearly marked, the trained strategy may overturn in the real world, so we must be cautious about this.

Alexander_Green_75

—

Our film and television preview team has been waiting for this for a long time. The traditional storyboard rendering cycle can take several days. If MoWorld can stabilize the lens control and background drift, the efficiency of the early proof of concept can be doubled several times. However, the online service has not been opened yet. We have to wait for the weights to be open sourced to measure the timing stability under long lenses. Most of the current demonstrations are at the clip level.

Judith_Hughes

—

To be honest, I have reservations about "localized narratives." At present, the weights and codes are not open source, and the real usability has not been verified at all. All data comes from official and media releases. Wait until GitHub releases reproducible training scripts and inference demos to judge whether it can really advance the world model to the industrial level. It is too early to draw conclusions now.

Kimberly.Parker78

—

I am looking forward to the direction of cloud gaming. Users can use W/A/S/D and the mouse to roam in the generated world in real time. If the delay can keep up, the experience will be very different. But whether the resolution above 1080P can be stable at 50 frames, and whether object drift will occur during long-term continuous exploration, these can only be known by actually playing it. The promotional video is too idealistic.

SOtho

—

Several reviews have said that it advances the world model from "capable of generation" to "capable of interaction and deployment". I agree with this positioning.

JudithRogersJr

—

50 frames! Now the world model is finally playable.

SHhar

—

After reading the technical report, 14B MoE is based on Wan2.2-A14B. Distillation reduces denoising from 50 steps to 4 steps, plus three-layer optimization of pipeline, parallelism, and operator. This set of projects is indeed solid. However, no matter how good the paper benchmark is, it is only laboratory data. If it is really going to be put into production, it depends on the end-to-end delay and service stability. We will wait for open source.

猫咪388

—

No one has mentioned the copyright boundary yet. Using real-world video training to reproduce interactive scenes will inevitably lead to commercial licensing pitfalls sooner or later.

Andrea.Barnes007

—

Don’t get carried away when people say the cost will be cut by 70%. It depends on the actual bill per unit time. It’s best to wait for the public beta to run and see the real price.

uwv3e

—

Compared with the 24 frames of Tencent HYWorld 1.5, the 50 frames of MoWorld are indeed a long way ahead. It depends on whether the community can reproduce it after open source.

KathleenMoore6

—

The fact that the cost can be reduced by 70% is the real ticket for small and medium-sized teams. It doesn’t matter if the big companies have their own computing power.

TylerHicks168

—

There is no problem with the idea of using it as a digital training ground for embodied intelligence training. Low cost and high fidelity are the key selling points.

Denise.SullivanII

—

50FPS, robot decision-making is saved.

DavidMurphy_Max

—

The purely domestic NPU can run through the real-time world model, and Shengteng is really up to speed, and there is no need to be stuck by NVIDIA.

RRoberts_7795

—

A post-2000 post-PhD student from Zhejiang University leads the team and has something.

LGMD8

—

DeepSeek moment for world models, for crying out loud.

Jennifer.Murray_88

—

A disciple of Academician Pan Yunhe started his own business, and his academic foundation is really strong. He has accumulated more than ten years of experience in 3D vision and spatial computing.

HashDex_dev

—

NVIDIA content 0, this title is too good.

RMitchell_2021

—

Genie replacement? Let’s look at open source first.

MGarcia_Max7

—

Wait for the open source and squat down.

CosmosCoder246

—

Nearly 1 billion in Series A, capital reached a consensus before the industry.

MoWorld

In-depth Report

User Reviews

Similar Products