Waypoint-1 is a real-time interactive video diffusion model developed by Overworld. It enables users to create immersive worlds through text, mouse, and keyboard inputs. The model is built on a frame-causal rectified flow transformer, trained on 10,000 hours of diverse video game footage along with control inputs and text captions. Unlike existing models, which often suffer from latency and limited control options, Waypoint-1 allows for fluid movement and interaction without lag. Users can manipulate the environment in real time, generating frames that respond immediately to their actions. Training involved a method called diffusion forcing, where the model learns to remove noise from future frames based on past ones. A causal attention mask ensures that each frame can only reference earlier frames, avoiding future context. However, the initial training method led to issues with error accumulation during longer video rollouts. To resolve this, the developers employed a technique called self-forcing, which aligns the training process more closely with how the model operates during inference, resulting in smoother outputs. The WorldEngine is Overworld's high-performance library for running Waypoint-1. It’s optimized for low latency and high throughput, allowing developers to create interactive applications using Python. The engine achieves impressive performance metrics, sustaining about 30,000 token-passes per second and generating frames at 30 FPS with four steps or 60 FPS with two steps. Key optimizations include AdaLN feature caching and matmul fusion, enhancing efficiency. Overworld is also hosting a hackathon on January 20, 2026, inviting teams to build upon the WorldEngine. The event offers a chance for participants to win a high-end GPU and network with others in the tech community.