In recent years, video generation models have advanced the development of "interactive video worlds" toward prototypes, yet they struggle to explicitly maintain high-level semantics such as objectives and rules. To address this, the AGI Lab at Westlake University and the team from Nanyang Technological University have proposed the Code World Model. This model centers around a language model, enabling a coding agent to maintain an executable world state by writing, invoking, and modifying code. These states are then converted into visual conditions through a proxy, guiding the video model to generate visuals. This design achieves a clear division of responsibilities where "code determines what happens in the world, and the video model determines its appearance." The Code World Model divides world operation into three parts: sparse decision-making by the coding agent, dense state updates through code execution, and visual observation generation by the video model. The model is trained using alignment data obtained from game runtime recordings. Prototype experiments demonstrate that it can adhere to the spatiotemporal constraints of the proxy, generating rich and diverse visuals, offering a new approach for constructing open-generated worlds.