World models become AI's next battleground, but deployment still faces three hurdles

Embodied intelligence, as represented by physical AI, has been one of the hottest terms in the tech world this year, and quite a few physical AI systems have already started entering factories. But the CEO of one embodied intelligence company once told us that most of the embodied intelligence entering factories today is not actually embodied intelligence at all; it is merely adding some AI capabilities to existing industrial automation equipment. True embodied intelligence is the ability to think independently to a certain degree across different production lines and different environments, and to perceive the world on its own. "Most robots today do not yet possess this capability, so they cannot be called embodied intelligence," the CEO said.

It is hardly a coincidence that at WAIC 2026, many industry experts also put forward the view that the core proposition of AI is shifting from generation to understanding and interaction. And for embodied intelligence to possess the ability to understand and interact, one indispensable technological foundation is the world model.

The world model is going through a critical stage. Leaping forward is at once the leap toward truly solving the problem of "understanding" the world and "simulating" the world, and the leap toward large-scale application in specific production and daily-life scenarios.

In this tough battle, who is charging at the very front?

Why do we need world models?

What exactly is a world model? Why have scientists of the caliber of Fei-Fei Li and Yann LeCun all placed their bets on it? And why did AMD not hesitate to spend heavily to acquire Fei-Fei Li's world model company World Labs?

To understand the value of world models, one must first return to the boundaries of large language models. A large language model is essentially "predicting the next word." By processing massive amounts of text, image, and video information, it answers what the next character, word, or sentence is. This mechanism performs impressively in dialogue, writing, and programming, but when AI moves from the screen into the physical world—when a robot has to grasp a cup, an autonomous vehicle has to drive through a rainstorm, a factory has to retrofit a production line—merely recognizing and generating content is far from enough. The system also needs to understand: where an object is located, how space changes, and what will happen to the environment after a certain action is taken.

What the world model fills in is precisely this layer. Its core value is to build an internal model of the real world: not only to understand the current state, but also to predict future states, simulate the outcomes of different actions, and provide a basis for planning and decision-making for robots or other intelligent systems. In repeated public statements, Fei-Fei Li has stressed one judgment again and again: if the real world is three-dimensional, machines must explicitly understand three-dimensional structure. Turing Award winner Yann LeCun's JEPA series of models takes a different approach, advocating prediction within an abstract representation space, which is in turn defined as an "implicit" technical route.

Explorers around the world are advancing toward the same goal along different paths. World Labs, founded by Fei-Fei Li, released its next-generation model Atlas in September 2026, billed as the world's first multimodal world model, with native support for text, images, video, and 3D data, capable of generating, reconstructing, and simulating three-dimensional worlds within a unified spatial context.

At the end of September, AMD acquired World Labs in an all-stock deal worth about US$8.2 billion, and Fei-Fei Li will serve as AMD's executive vice president and chief scientist. At the same time, Nvidia acquired the open-source model community Hugging Face for about US$12.9 billion. The two chip giants reached into the model layer one after the other in the same month.

The domestic market is equally lively. From the computing power foundation to general-purpose foundation models and then to vertical models, this bears similarities to the industrial landscape of large language models. The Beijing Academy of Artificial Intelligence released Wujie·Physis-v0.1, focusing on unified physical state modeling, while leading domestic tech companies such as Kunlun Wanwei, Tencent, and Alibaba have also made their own arrangements.

But the industrial and sectoral significance of world models goes far beyond being "the next big large-model trend." In fields such as advanced manufacturing, embodied robots, autonomous driving, and scene exploration, internet data can provide only very weak capabilities for understanding and applying the physical world that AI agents situated in the physical world require. What is truly scarce is data carrying information about physical changes such as gravity, friction, materials, and speed; interactive, verifiable, and reusable simulation environments; and closed-loop capabilities from training to deployment.

The value of world models lies precisely in advancing generated content into interactive, verifiable, and reusable spatial assets—for example, generating extreme scenarios in autonomous driving to reduce the cost of real-vehicle road testing; rehearsing actions and validating decisions in embodied intelligence to shorten the generalization cycle across production lines and environments; supporting production line retrofit rehearsal, risk identification, and remote inspection in industrial digital twins; and enabling 3D/4D assets to be repeatedly rendered, edited, and invoked in film and television, gaming, and consumer devices.

This also means that the logic of industry competition will shift from simply selling model interfaces to providing integrated services spanning models, computing power, data, simulation environments, and industry systems. Therefore, the integrated software-hardware delivery of "chips + models" is regarded as a prominent feature of world models once they enter industrial application, and chip modules have accordingly become an important part of the industrial layout.

Moxin Technology is a typical sample within this industrial system. Chen Tianrun, founder of this Hangzhou-based company, focused on 4D spatial modeling during his academic research, and since the end of 2024 the company has also devoted itself fully to the spatial intelligence track:

In December 2025, Moxin Technology released the KOKONI-World prototype, an interactive video world model. In July 2026, Moxin Technology, together with the National Artificial Intelligence Application Pilot Base and Huawei Cloud, released MoWorld 1.0, the first ultra-high-frame-rate interactive world model, and proposed the concept of the "Flash World Model," marking AI's shift from generating content toward interactive world simulation.

And at the World Internet Conference held just recently, Moxin Technology, together with the National Artificial Intelligence Application Pilot Bases in the three directions of embodied intelligence, cultural tourism, and spatiotemporal geographic information, released MoWorld 2.0, a 4D world model built natively on Huawei Ascend. According to the introduction, the model can deeply understand more than a dozen physical elements such as gravity, elasticity, buoyancy, lighting, soft bodies, fluids, rigid bodies, and heat, and integrates them into its overall model capabilities; it can reason about how the world changes after an action occurs and output 4D Gaussian assets that can be generated by simulation, rendered in real time, and repeatedly invoked. Its real-time inference speed has increased from a maximum of about 50 FPS in version 1.0 to about 70 FPS, 3D/4D asset rendering reaches over 500 FPS, and on mobile devices it is about 100 FPS. The model is trained and deployed on fully domestic Ascend computing power, and inference costs can be reduced to as low as 30% of GPU solutions of the same scale; it can be used in scenarios such as embodied training, autonomous driving, industrial digital twins, film and gaming, and cultural tourism.

This pace of technological iteration has attracted enormous attention from the capital market. Over the past 10 months, Moxin Technology has completed 5 consecutive rounds of financing, raising a cumulative 2 billion yuan. Investors include state-owned capital such as Shenzhen Capital Group and Zhejiang Venture Capital, industrial capital such as Legend Holdings, Transsion Technology, China Telecom, Huace Film & TV, Fullhan Microelectronics, and SMIC Capital, as well as multiple technology VCs. The investor structure covers chips, computing power, intelligent terminals, digital content, and advanced manufacturing.

Some analysts believe this means that the multi-party layout for the industrialization of world models has entered a critical stage, and core players across the upstream and downstream of the industrial chain are collectively betting on the industrialization value of world models.

Of course, these investment institutions are optimistic not only about a single vendor making world models, but also about the world model track behind Moxin Technology. According to Morgan Stanley's forecast, by 2035 the scale of industries empowered by world models will reach 10 trillion US dollars.

The Three Hurdles Behind World Models

Behind the bustling financing and product launches, the world model industry is in fact still at an early stage. Almost all practitioners admit that there are at least three hurdles ahead.

The first hurdle is the cost of acquiring real-world data. This issue is often underestimated. The text data required for training large language models is already highly digitized, and the cost of acquisition and annotation is relatively controllable. But world models need more than video and text; they need corresponding information about physical changes: gravity, friction, materials, speed, and so on.

High-quality world model training data needs to simultaneously describe visual appearance, camera motion, spatial structure, and dynamic changes. According to the introduction, based on its accumulation in 3D/4D visual geometry, Moxin Technology has built a geometry-aware data engine that filters and governs training data through camera pose estimation, 3D reconstruction, dynamic region filtering, trajectory verification, and multi-view consistency evaluation, forming a continuous data loop through data collection, model training, scenario validation, and application feedback. This engineering capability is not sexy, but it determines the lower bound of the model.

The second hurdle is the difficult-to-reconcile contradiction between real-time performance and cost. This is a hard constraint that distinguishes world models from large language models. When a large language model generates a reply, users can still accept a delay of a few seconds; but world models usually have large parameter counts and high computational loads. If they cannot respond quickly, closed-loop systems such as robots and autonomous driving will find them difficult to use—a robot cannot wait two seconds to learn that there is a step ahead; and if inference costs are too high, large-scale deployment becomes difficult.

The other side of the contradiction is hardware. At present, most mainstream world models rely on high-end GPUs, which is precisely why Moxin Technology chose the domestic NPU route. According to company disclosures, the MoWorld series models have been optimized across the entire chain from data construction and model training to system deployment, laying the foundation for large-scale deployment on devices.

The third hurdle is the consistency gap between the generated world and the real world. A model may generate content that looks reasonable, but whether it can achieve consistency with the real physical world in physical laws, collision relationships, object states, and causality still needs verification. World models are moving from generative simulation tools toward interactive training environments, but before they can directly assume full control responsibilities in high-risk scenarios, more real data, physical validation, and industry testing are needed. From looking real to being behaviorally reliable, and then to being industrially usable, this road is quite long. And the ability of Moxin Technology's MoWorld 2.0 to integrate and understand multiple physical elements is precisely one of the key technological leaps.

Where Do World Models Land?

Despite the many challenges, the industrialization of world models has already run ahead in several directions. The first is some consumer-level scenarios with relatively low requirements. For example, Moxin Technology, together with Huawei, released the HarmonyOS application "Moxiaochong."

Users only need to upload a few photos of their pet to generate, right on their phone, a 4D digital pet avatar with spatial structure that can interact and continuously change. Behind this is Moxin Technology's MoWorld 2.0 model, which not only completed training on the Ascend platform but was also quantized and pruned to form an on-device version, running a model of roughly 600 million parameters on Kirin chips. This is the first time a 4D world model has entered a mobile application, and it also means world models are beginning to move from the cloud into end users' pockets.

Beyond consumer applications, the value of world models is also gradually emerging in B-end industrial scenarios. For example, in the autonomous driving collaboration between Pony.ai and Moore Threads, the world model can generate more than 10 billion kilometers of test data per week and can derive hundreds or even thousands of variations of "hazardous situations," allowing the virtual driver to continuously evolve through repeated training. Bao Shiqiang, CEO of 51World, pointed out that what truly constrains the continuous evolution of models is not the total volume of data but the ability to supply high-value data; simulation is an amplifier of high-value data. In embodied intelligence training, the ω-EVA world model released by Xingyuanzhi achieved a decision-making closed loop of "rehearsal, validation, action" for the first time, with an average success rate of 98.6% in the LIBERO test. In the field of industrial digital twins, the pipeline network industrial world model exhibited by PipeChina is being used to support simulation deduction under complex operating conditions and intelligent identification of potential risks.

The general-purpose world model technology strategy route chosen by Moxin Technology has given it a foundation for a broader layout in B-end applications. For example, in the field of industrial digital twins, the MoWorld 2.0 model is being used to transform real environments such as factory production lines and mines into spatial assets, serving equipment layout, renovation rehearsal, training collaboration, and remote inspection; in the field of embodied intelligence training, MoWorld 2.0 can build virtual training environments with spatial structure and basic physical feedback, supporting robots in repeated training and trial and error; in the field of autonomous driving training, the model can generate long-tail scenarios such as rain and snow weather, potholed roads, sudden obstacles, and complex traffic environments in a unified 3D/4D space, improving training efficiency; in the fields of urban and rural planning and cultural tourism, the model can be used for spatial scheme presentation, virtual tours, digital scenic areas, and immersive content production; in the fields of film and television and gaming, MoWorld 2.0 can quickly generate reusable scene assets, support free camera movement, object editing, and weather switching, greatly compressing the cycles of early-stage set construction, storyboard rehearsal, and concept validation, and improving content production efficiency......

As world models continue to penetrate various application scenarios, future business models will also change, and the focus of competition will shift from simply selling model interfaces to providing integrated services of models, computing power, data, simulation environments, and industry systems. Whether a model can run continuously, whether it has controllable costs, and whether it can connect to customers' existing systems will be as important as model performance.

Facing this trend, Moxin Technology's combination is "general-purpose 4D world model + domestic computing power foundation + industrial scenarios," and it hopes to drive world models from laboratory demonstrations to real applications. "World models will ultimately not just be a standalone product, but may become infrastructure commonly used by robots, intelligent vehicles, industrial software, digital twins, and new content platforms." This is the prediction of Chen Tianrun, founder of Moxin Technology.