Chinese Team Achieves Architectural Breakthrough in World Models
Recently, InSpatio, a startup team from Zhejiang University, has officially released and open-sourced its world model, InSpatio-World. On the authoritative benchmark WorldScore-Dynamic, its overall performance ranks first among models capable of real-time and interactive-level inference speeds. Notably, this achievement was realized with a total training cost in the hundreds of thousands of dollars and a compute scale of around one hundred GPUs, showcasing the potential of algorithmic architecture innovation in competing with sheer computational scale.
A world model is seen as the key for AI to understand and reconstruct the physical world. At its core is the construction of an internal “physics engine” within the AI, used to comprehend 3D space, remember the states of objects, and predict their dynamic evolution. InSpatio-World can transform ordinary monocular video data into a real-time, interactive, dynamic 4D space. It achieves an inference speed of 10 FPS on a single NVIDIA RTX 4090 GPU and can reach 24 FPS on professional-grade GPUs, significantly lowering the barrier to digitizing the physical world.
Diverging Technical Paths: From 2D Visual Statistics to 3D Spatial Modeling
Currently, there are two main technical paths for building world models. The first is a 2D video-based approach, such as NVIDIA’s Cosmos and Google’s Genie. Its advantage lies in leveraging massive amounts of video data for learning, but it faces challenges in handling complex occlusions or maintaining long-term physical consistency. The second is a 3D space-based approach, like World Labs’ Marble. This path inherently possesses spatial consistency but is limited by the scarcity of high-quality 3D data.
InSpatio chose the more challenging but more first-principles-aligned 3D path. Its core technical innovation is the proposal of the “State-Anchored World Modeling” paradigm. By explicitly modeling the “world state,” it decouples the observation viewpoint from the physical entities. This allows temporal evolution to be represented as physical updates to a 3D state, rather than the continuous generation of 2D pixels. By combining three key techniques—“Explicit State Modeling,” “Spatio-temporal Autoregressive Framework,” and “Joint Distribution Matching Distillation”—the model effectively “distills” the visual realism of massive 2D video data into 3D space, thereby achieving a leap from “pixel simulation” to “physics simulation.”
The direct result of this paradigm shift is that the model no longer generates linear video clips, but an interactive and retrospective 4D spacetime. Users can freely pause, rewind to any historical moment, and interact with the dynamic digital world, ensuring the consistency of physical logic over long-term evolution.
Algorithm and Engineering Innovation Driven by Industry-Academia-Research Collaboration
InSpatio’s successful breakthrough benefits from its unique path of industry-academia-research collaboration. The team’s Chief Scientist is Professor Bao Hujun from Zhejiang University, whose team has deep expertise in fields like computer graphics and spatial computing, providing the theoretical foundation for the model. Founder Professor Zhang Guofeng, also from Zhejiang University, gained extensive industry experience at SenseTime, effectively bridging the gap between academic frontiers and industrial application.
Co-founder Dr. Liu Haomin and his engineering team are responsible for the engineering implementation and efficiency optimization of cutting-edge algorithms. This closed loop, from theoretical breakthrough to engineering validation, allows the team to transform complex spatial intelligence algorithms into high-performance, real-time systems with limited computing resources. InSpatio’s successive releases of the static 3D reconstruction model InSpatio-WorldFM and the dynamic 4D world model InSpatio-World in a short period are a testament to the success of this collaborative model.
World Models Poised to Reshape Applications Across Multiple Industries
World models are considered the next key technology after Large Language Models (LLMs) that could define the next wave of AI. It moves AI beyond content generation towards a deep understanding, prediction, and planning of the physical world, with its influence spanning several key industries.
- Autonomous Driving: World models can be used for generative simulation to create and test a vast number of corner cases in a virtual world, forming a closed loop of “data-simulation-policy iteration” and driving the industry from perception-based to generative intelligent driving.
- Embodied Intelligence: Robots can use world models to perform internal reasoning and simulation before executing tasks, predicting the consequences of different decisions, thus transitioning from passive execution to proactive planning.
- Virtual and Augmented Reality (VR/AR): This technology is expected to solve the production bottleneck for spatial content by automatically generating explorable 3D environments from multimodal data such as text and images, reducing the cost of content creation.
The InSpatio team stated that they have already received business collaboration inquiries from dozens of global companies in the robotics, autonomous driving, and content creation sectors, and are accelerating the commercialization and industrial application of InSpatio-World.
