Chinese World Model Tops WorldScore

Homegrown World Model Takes First Place on WorldScore

As of July 24, the top position on WorldScore, the world model benchmark leaderboard launched by Fei-Fei Li’s team, has changed hands.

The new No. 1 is UniWorld-View, jointly developed by the Tuzhan Intelligence team, Peking University, and Peng Cheng Laboratory. Its core capability can be summarized in one sentence: given a single image, the model does more than fill in a few additional angles beyond the visible frame—it attempts to extrapolate a complete scene that can continue to be observed from different viewpoints.

More importantly for developers, the team did not release only demo videos and a technical report. It also made the model code and weights available and completed adaptation for Huawei Ascend computing hardware. In other words, this is not a demo that can only be viewed on a paper’s webpage. Developers can download, deploy, and validate it, with the possibility of continuing training or performing industry-specific fine-tuning on domestic computing clusters.

If viewed only as “a Chinese model topping a leaderboard created by an overseas team,” the news could easily be interpreted as yet another benchmark victory. But UniWorld-View’s real value lies in the combination of three factors: world-generation capabilities, full open sourcing, and adaptation to domestic computing hardware.

None of these is novel on its own. Accomplishing all three at once is what gives the model the potential to enter real-world R&D workflows.

It Generates More Than an Image—it Generates a Scene Hypothesis

A world model is neither a larger text-to-image model nor a large language model with a visual interface bolted on.

A conventional image-generation model addresses the question, “Does this frame look real?” A world model must go a step further and answer, “If the observer turns around, walks forward, or moves behind an object, what should appear there?”

The difference between these two tasks is much like the difference between drawing a movie poster and building a game map.

A poster only needs to look good from the front; there may be nothing behind it. A game map, however, must maintain spatial continuity: What is behind the door? What shape does the building have from the side? How do near and distant elements correspond? Do objects remain in their original positions after the camera moves? Even if every frame generated by a model looks beautiful, it has not truly established a stable world representation if a table disappears when viewed from another angle or a room fails to connect back to itself after the camera makes a full circuit.

UniWorld-View takes a very direct—and very difficult—approach: extrapolating other views of a scene from a single image. The information provided by one image is inherently incomplete. A model may see the front of a building but have no way of knowing whether there are windows on the back. If a sofa blocks a corner of the wall, the model cannot directly recover the occluded area from the visible pixels.

This task is therefore not simply about “recovering the truth.” It is about generating a scene hypothesis that is plausible, continuous, and as consistent as possible based on the available visual evidence.

The process can be abstracted as follows:

Single reference image
    ↓
Extract objects, layout, depth, and scene semantics
    ↓
Build an implicit spatial representation
    ↓
Extrapolate unobserved regions along the target viewpoint
    ↓
Generate scene images with cross-view consistency

The hardest part is not generating one new viewpoint, but ensuring that multiple new viewpoints do not contradict one another. For a world model, “every image looks good” is only a baseline requirement. The core criterion is whether “these images appear to come from the same place.”

What Topping the Leaderboard Means—and What It Does Not

The significance of WorldScore lies in its attempt to evaluate world-generation models under a standardized protocol rather than merely comparing demo clips handpicked by vendors.

World models are particularly prone to creating misleading demos. When only a predetermined camera path is shown, teams can avoid difficult cases such as complex occlusions, rapid rotations, long-distance movement, and scene loops. A leaderboard at least places different models within a relatively consistent evaluation framework, giving developers an opportunity to compare their performance on similar tasks.

UniWorld-View’s top ranking indicates that it has achieved leading results on the tasks and evaluation system currently covered by WorldScore. It also demonstrates that Chinese teams have entered the top tier in spatial generation and viewpoint consistency.

However, the limits must be stated clearly: ranking first on a leaderboard does not mean the world-model problem has been solved.

First, every benchmark has its own task distribution. A model may excel at natural scenes, indoor spaces, or common camera movements, yet still struggle with industrial equipment, dense urban roads, or highly dynamic environments. Leading on a test set does not automatically translate into stability in production environments.

Second, visual plausibility does not equal physical correctness. A model may generate a street that looks coherent without understanding vehicle collisions, gravity, object permanence, or causality. In the strict sense, a world model must do more than “see the world”; it must also predict how actions will change it.

Finally, short-range viewpoint expansion and long-term state preservation are two different things. Remaining error-free during a few seconds of camera movement does not mean the model will still remember the original spatial layout after several minutes of continuous interaction.

A more accurate assessment of this achievement is therefore: UniWorld-View offers a highly competitive current solution for single-image-driven world-view generation, but it remains a key piece of the puzzle on the path toward interactive, predictive world models—not the final destination.

Fully Open Code and Weights Matter More Than First Place

For developers, a leaderboard ranking provides a reason to pay attention; the availability of code and weights determines whether a model can have a lasting impact.

In the past, many world-model projects adopted a release strategy of “public paper, private weights, limited demo access.” The reasons are understandable: training is expensive, data processing is complex, and video and spatial generation models can be used to create deceptive content. The problem is that without weights, results are difficult to reproduce, and there is no way to determine whether a model truly possesses general capabilities or has merely been heavily optimized for its showcase examples.

By opening its code and weights, UniWorld-View delivers at least four direct benefits:

• Reproducibility: Researchers can verify the leaderboard results using their own data and hardware rather than simply trusting official videos.
• Fine-tunability: Automotive, gaming, film, and robotics teams can perform targeted training with industry-specific data.
• Inspectability: Developers can analyze its spatial representation, viewpoint control, and generation modules instead of facing a black-box API.
• Deployability: Enterprises can evaluate the cost, latency, and throughput of private deployments without uploading sensitive images to third-party platforms.

However, “fully open-source code and weights” should not be taken to mean unrestricted commercial use. Before integrating the model, developers still need to review the repository’s license, training-data documentation, derivative-model terms, and restrictions on generated content. Being downloadable, being runnable, and being usable in commercial products are three different things.

Open-source models also offer an often-overlooked advantage: negative results become visible as well. The community can more quickly identify which camera paths cause scene drift, which objects suffer structural collapse, and how to balance VRAM usage, throughput, and quality. For a technology such as world models that is still undergoing rapid exploration, publicly documented failure cases often have greater research value than another ten carefully selected videos.

Ascend Adaptation Is More Than a Slogan About “Domestic Technology”

UniWorld-View’s adaptation to Ascend computing hardware is the most clearly valuable engineering aspect of this release.

Most current research on generative models still assumes the use of NVIDIA GPUs and the CUDA ecosystem. Even if model weights are released, Chinese enterprises may still receive little more than “open source on paper” if key operators, precision strategies, parallelization schemes, and runtimes are available only on CUDA and cannot readily be deployed at scale.

Migrating to Ascend involves far more than changing the device name from GPU to NPU. World-generation models typically include high-resolution visual encoding, diffusion or video-generation modules, and potentially large amounts of spatiotemporal attention computation. Genuine adaptation must address at least the following questions:

1. Can all operators required by the model run properly within the Ascend software stack?
2. Does mixed-precision execution cause significant losses in visual quality or consistency?
3. Are memory management and long-sequence computation stable?
4. Is communication overhead during multi-card inference and training manageable?
5. Can equivalent alternatives be found for custom optimizations in the original CUDA implementation?

“Adapted” is therefore a starting point, not a conclusion about performance. What matters next is whether the team provides reproducible Ascend deployment instructions, along with measurements of speed, memory usage, and scaling efficiency under identical precision, resolution, and sampling configurations.

Still, it resolves the fundamental question of availability: developers do not have to migrate the entire model from scratch. For government, industrial, transportation, and research organizations that require on-premises deployment, this is far more useful than a cloud-based demo.

In the past, adaptation to domestic computing hardware often occurred months after a model’s release and was concentrated mainly on language models. UniWorld-View included Ascend support at launch, indicating that China’s domestic computing ecosystem is expanding into more complex workloads such as video generation and spatial intelligence.

Who Will Use World Models First?

World models are most often discussed within the grand narratives of robotics and autonomous driving, but in the near term, their most practical applications may not involve directly controlling machines. Instead, they may be used to generate training data and content assets.

Games and Virtual Environments

Art teams can start from concept images and rapidly explore different camera positions within the same scene for level previsualization, skybox generation, or rough scene prototyping. This will not immediately replace 3D modeling because the generated results may lack editable meshes and strict geometric structure, but it can significantly reduce the time required for early-stage visual exploration.

Film and Advertising

Directors or storyboard artists can input a scene image and preview its approximate appearance under dolly, orbiting, and lateral camera movements. What is needed here is a visual proposal rather than millimeter-level spatial accuracy, making these applications relatively tolerant of some of the model’s current shortcomings.

Autonomous-Driving Simulation

Real-world road data is expensive, and extreme weather, accidents, and rare combinations of traffic conditions are especially difficult to collect. World models can supplement long-tail scenarios, but only if the generated data not only looks real but also preserves the states of road users and the consistency of road structures. Otherwise, training systems on data generated according to incorrect physical rules may be more dangerous than having insufficient data.

Robot Training

Robot policies need to understand the relationship between actions and environmental changes. View-generation models such as UniWorld-View can provide spatial priors, but to serve as robot-training environments, they must also incorporate action conditioning, state transitions, collision feedback, and long-term memory.

This is also the dividing line for the next stage of world models: moving from “what the world looks like after the camera moves” to “what happens to the world after an action is performed.” The former focuses on spatial generation; the latter is much closer to the environmental simulators required for embodied intelligence.

What Homegrown World Models Really Need Are Better Engineering Metrics

UniWorld-View’s top ranking deserves recognition, but competition among world models will not stop at a single leaderboard result.

Going forward, developers should pay closer attention to several metrics that rarely appear in promotional videos:

• Does the spatial structure gradually drift as continuous generation time increases?
• When the camera leaves an area and later returns, do the original objects remain intact?
• Does the model fail noticeably on out-of-distribution industry images?
• How much VRAM, time, and how many sampling steps does each generation require?
• Do the Ascend and GPU versions deliver consistent quality and throughput?
• Can the model’s output be converted into downstream-ready representations such as depth maps, point clouds, and meshes?
• How much data and computing power are required for fine-tuning?

These metrics will determine whether UniWorld-View is merely a strong research result or a foundation model that can genuinely enter production environments.

Our assessment is that the most commendable aspect of UniWorld-View is not that it is “China’s No. 1,” but that it combines leading benchmark results, open weights, and deployment on domestic computing hardware in a single release. It has not yet demonstrated a complete understanding of physics, nor should its first-place ranking exempt it from real-world testing. But for developers interested in spatial intelligence, generative simulation, and controllable viewpoint generation, it is already an open-source baseline worth downloading and evaluating.

Open-source competition among large language models has already progressed from “Are the weights available?” to “Can the model be deployed at low cost?” World models are now following the same path. UniWorld-View has taken a step forward: it not only generates a world, but also gives outside developers the opportunity to enter that world and examine how solidly it has been built.

References

• Zhihu: Homegrown World Model Tops Fei-Fei Li Team’s Leaderboard, Adapted for Domestic Ascend Computing Hardware—An introduction to UniWorld-View’s top WorldScore ranking, open-source release, and Ascend adaptation.