University of Beijing launches 5D world model that evolves – EvoPhys-World

The University of Beijing has just changed the rules of the game to create synthetic worlds. The new EvoPhys-World is described as the first world model 5D fully controllable and at scene level, going far beyond traditional 3D or 4D models. Built on the domestic computing infrastructure of Chinese Moore Threads, the model allows AI-generated environments to evolve and interact with the user in real time — without relying on foreign hardware or external cloud.

5D World Model: What does this new dimension mean?

While common models work with width, height, depth and time (the four classical dimensions), the EvoPhys-World adds a fifth dimension: the controlled and continuous interaction capacity with human agents and intelligent objects. In practice, this means that each element of the scene can be physically simulated — collision, deformation, response to commands — and the world itself “learns” with the actions of the user, adjusting its structure in real time. It's a jump from 'watchable and navigable' to 'fully interactive and self-evolutionary'.

Differences of the World Model 5D Self-Evolutionary of the University of Beijing

Previous models, such as those used in robotic simulations or video games, are usually pre-programmed or limited to predictable responses. The EvoPhys-World operates in a continuous loop of perception-action-reaction. With each user interaction, the model recalculates the physics of the scene, the position of objects and even the illumination — all processed locally, without server latency. This allows, for example, an architect to test modifications in a virtual building and instantly see how the structure behaves under different loads, without needing supercomputers.

Moore Threads and home infrastructure in model 5D

All the computational power of EvoPhys-World comes from the infrastructure of Moore Threads, a Chinese company that develops GPUs and domestic computing solutions for AI. When running locally, the model eliminates dependency on foreign servers and accelerates response time for fractions of a second. This is crucial for real-time applications such as medical simulations, robot training and immersive environments. It is no coincidence that the announcement takes place amid the [million-dollar race for trillion-dollar valuations in the IA sector] (/noticias/corrida-bilionaria-empresas-de-ia-fila-ipo-valuations-trilhoes-bolha) — computational sovereignty has become a strategic asset.

Impact of the 5D world model for Brazilian developers

With EvoPhys-World running on affordable hardware, Brazilian startups can create hyper-realistic physical simulators without having to rent cloud clusters. Imagine testing a drone in a 5D environment before building it, or training computer vision algorithms in scenes that modify themselves. In addition, the open-weight (though partial) model approach allows local teams to adapt it to scenarios such as precision agriculture or increased retail reality. The entrance barrier just went down.

Moore's domestic infrastructure Threads in 5D World Model

It refers to chips and graphics cards designed and manufactured in China by Moore Threads, which offer competitive performance with NVIDIA GPUs for AI workloads and simulation. By using this base, EvoPhys-World ensures that all processing is on the local hardware without external communication. This not only reduces latency, but also increases data security and privacy — a differential for sectors such as health and defense.

Frequently Asked Questions about the 5D World Model EvoPhys-World

Is the model available for use?

There is no official public release date yet, but researchers can request early access through the University of Beijing website. The expectation is that a limited version will be released for testing in the third quarter of 2026.

Which languages or frameworks are compatible?

EvoPhys-World was developed in Python with PyTorch, and Moore Threads provides drivers and SDKs compatible with major deep learning libraries. Integration with Unreal Engine and Unity is in beta phase.

Does this model replace the traditional metaverse?

Yes, in applications where physical interaction and environmental evolution are crucial. EvoPhys-World goes beyond static or pre-rendered worlds, offering an ecosystem that responds organically to the user—which the current metaverse still does not deliver.

> Source: Pandaily – Peking University EvoPhys-World 5D World Model