World models have become a new focal point for AI developers trying to give software a more durable understanding of space, movement, and physical interaction. QbitAI reported on September 4 that World Labs, the company founded by Fei-Fei Li, released its multimodal world model Atlas on September 2. The report also pointed to a Chinese response: InSpatio, a spatial-intelligence team that open-sourced InSpatio-World in March, has topped a new benchmark and launched an open 3D-data initiative.
The comparison should not be overstated. Atlas and InSpatio-World are not identical systems, and QbitAI does not provide a head-to-head evaluation. The story shows that Chinese teams are building their own world-model stack, combining model releases, benchmarks, data plans, and industry partnerships rather than waiting for a foreign platform to define the field.
QbitAI said Atlas brings text, images, video, and 3D information into a common spatial context. The system is designed to reconstruct scenes from one or more images, create new viewpoints, and move a camera through a generated environment while preserving structure. Those capabilities put Atlas in the growing category of systems meant to model a space rather than merely generate a plausible video clip.
InSpatio’s Benchmark Result Focuses on Physical Consistency
The strongest current data point for InSpatio is its result in WorldArena 2.0. QbitAI reported that its InSpatio-Curious model ranked first among 77 models in Track 1 of the benchmark as of August 27, with an overall score of 66.11. It listed scores of 73.24 for Physics Adherence, 64.89 for Trajectory Accuracy, 98.36 for JEPA Similarity, and 99.29 for Depth Accuracy.
Those categories matter because world models are increasingly evaluated on more than visual polish. A video may look realistic but still fail to preserve the trajectory of an object, the depth between two surfaces, or the likely result of a physical interaction. QbitAI describes WorldArena 2.0 as a benchmark that examines standardized simulation and robot-manipulation tasks, including motion trajectories, depth relations, and aspects of physical plausibility.
The same report notes an important limitation. InSpatio-Curious recorded an image-quality score of 60.64, nearly nine points below the second-place overall model, yet finished first on the combined ranking. That result suggests WorldArena is rewarding more than surface realism. It does not show that InSpatio has solved real-world robotics or that it will perform equally well in every factory, home, or outdoor setting.
Benchmark results are useful diagnostic tools, but they come from a fixed set of tasks and data distributions. QbitAI itself says WorldArena 2.0 remains centered on standardized simulation environments and manipulation tasks. It does not directly measure every physical force, material property, or long-tail event that a robot might encounter outside the test set.
InSpatio’s March open-source release gives the result a broader strategic meaning. The team’s model is designed to infer a dynamic 4D scene from video and allow users to change camera trajectory or viewpoint. That places it in the same wider technical movement as models that try to represent a persistent environment, not just produce one static output.
SIDO Tries to Build a Shared 3D Data Layer
InSpatio is pairing its benchmark result with a data initiative. QbitAI reported that it launched the Large-scale Spatial Intelligence 3D Data Open Plan, known as SIDO, at the second China Spatial Intelligence Conference in Wuhan on August 29. The source said more than 20 university teams and industry organizations participated.
SIDO’s stated purpose is to connect data producers, model builders, and application users around 3D and 4D spatial information. QbitAI says the plan covers universities, lidar providers, robotics companies, 3D-generation developers, digital-space companies, and industrial users. The participants aim to create a larger data foundation and develop benchmark work for 3D reconstruction and generation, scene understanding, spatial reasoning, dynamic prediction, and embodied planning.
The initiative is significant because spatial AI often lacks the standardized data that helped earlier computer-vision systems improve quickly. A world model may need video, geometry, depth, motion, object interaction, and task-specific labels. Those datasets can be expensive to gather and difficult to share across industries. An open plan can lower duplication, but it also has to resolve data quality, permissions, formats, and evaluation standards.
China already has several companies pursuing related tools. EastFrontier reported in August that Manycore launched Lux3D for China’s spatial-AI push. It also covered how Amap turned 12 frames into long-range 3D AI. These projects take different technical paths, but they share an interest in turning limited visual input into a usable three-dimensional representation.
Atlas Raises the Bar Without Settling the Race
Atlas gives World Labs a high-profile entry into the field, supported by Fei-Fei Li’s standing in computer vision and spatial intelligence. QbitAI says Atlas can take visual inputs and build an environment that can be viewed from new camera positions. InSpatio-World starts from a different premise, using video to create a dynamic scene where time as well as camera position can be varied.
The technical differences are more important than an easy national comparison. QbitAI characterizes Atlas as an image-centered approach to a navigable virtual environment and InSpatio-World as a video-centered approach that explicitly includes time. Each may work better for different tasks. A design tool may value fast reconstruction from still images, while a robot-training system may need coherent movement and interaction through time.
The China story is therefore not that a local model has conclusively displaced a U.S. model. It is that the release of Atlas arrives as Chinese teams are building their own open-source models, setting benchmark targets, and organizing data collaborations. InSpatio’s WorldArena result and SIDO launch provide measurable evidence of that effort, while still leaving open the question of how well these systems will transfer from tests to products.
The next test will be deployment. Spatial models have potential uses in robot training, industrial simulation, gaming, mapping, design, and digital twins. Developers will need to demonstrate that they reduce the cost of collecting data, improve planning, or make physical systems more reliable. The September developments show that China is treating world models as a full-stack competition involving data, benchmarks, and applications, not simply a race to release the most striking video demo.