The gulf between AI world models that produce slick demo reels and those you can actually interact with has been vast. Overworld just closed that gap with Waypoint-1, a real-time video diffusion model that doesn’t just generate video game footage—it lets you play it. Using text, mouse, and keyboard, you can step into a hallucinated world with zero latency, and it runs fast enough to hit 60 frames per second on consumer hardware.
The secret isn’t just a bigger model. It’s a fundamental rethink of the training and inference pipeline. Instead of taking a pre-trained video model and bolting on simple controls as an afterthought, Waypoint-1 was built from scratch for interactivity on 10,000 hours of diverse gameplay data. It uses a frame-causal rectified flow transformer. The team first pre-trained with “diffusion forcing,” teaching the model to denoise frames while respecting causality. They then fixed the classic autoregressive drift problem with a post-training technique they call “self forcing,” which uses distribution matching distillation (DMD) to align training with the actual frame-by-frame rollout used during inference.
That theoretical alignment translates to raw speed thanks to WorldEngine, Overworld’s custom inference library. On a single RTX 5090, the 2.3-billion-parameter Waypoint-1-Small model sustains roughly 30,000 token-passes per second. That translates to a playable 30 FPS at 4 denoising steps or a buttery 60 FPS at 2 steps. The performance comes from aggressive optimization: caching AdaLN features, fusing matrix multiplications, and a static rolling KV cache that avoids redundant computation. The result is a Python library where defining a new world is as simple as typing a prompt like “A game where you herd goats in a beautiful valley” and then looping over controller inputs.
Overworld isn’t keeping this locked in a lab. Weights for the small model are on Hugging Face right now, with a medium version on the way. They’re also putting a literal prize on the line to see what people build. The company is hosting an 8-hour hackathon on January 20th where the winning team walks away with a 5090 GPU. It’s a clear signal that they view this less as a finished product and more as the first viable platform for a new kind of software—one that is streamed from a model, not rendered by a traditional engine.
