Peking University Launches World's First Self-Evolving 5D World Model EvoPhys-World

A research team from Peking University has unveiled EvoPhys-World, described as the world's first human-centric, scene-level, fully controllable 5D world model. Built on Moore Threads' domestically-produced computing infrastructure, the model pushes AI-generated worlds from the "watchable and navigable" stage to one that is "manipulable, deeply interactive, and self-evolving." The EvoPhys team argues that the essence of a world model is the simulation of a hyperdimensional space. While 3D world models focus on spatial structure and 4D models add the dimension of time, a 5D world model must also understand the states of objects across different parallel universes — how different choices lead to different futures, and how those future simulations inform current decisions. EvoPhys-World is built on a unified base model with two core forms: Model as World Engine and Model as World Policy. In the World Engine form, the model can construct continuously evolving scene-level world memories based on real environments and generate future imaginations. It supports arbitrary trajectory roaming within new scenes and simulates interactions with any object in the scene, including manipulation and movement. In the World Policy form, the model can predict future world states and recommend optimal action strategies. This enables it to understand not just how the world evolves, but to pre-enact how an intelligent agent should act next, forming a complete loop from world generation to world manipulation. Technically, EvoPhys-World employs a Latent Memory Pool for long-term spatiotemporal memory, using a spatiotemporal importance mechanism to select and compress key implicit states. It also features a parallel generation architecture with Unified Token Chunk output and a hybrid attention mechanism that performs unified state-action token parallel inference. The underlying computing infrastructure is powered by Moore Threads' full-stack domestic GPU ecosystem, marking a significant milestone for China's homegrown AI computing capabilities. By relying entirely on domestically produced hardware, the project demonstrates that advanced world model research can be conducted without dependence on foreign GPU supply chains. The project is hosted at evophys.com and represents a significant step forward in the world model research paradigm — moving AI from "seeing the world" to "changing the world" through understanding physical causality, interaction dynamics, and multi-future prediction and selection.