Overworld · 2026-04-09 · major
Waypoint-1.5 — Interactive 3D World Model at 720p/60fps on Consumer GPUs
Overworld's Waypoint-1.5 generates interactive first-person 3D worlds at 720p/60fps on RTX 3090+ hardware (360p on gaming laptops). Trained on 100x more data than v1. Reviewers describe it as playable like an actual FPS — not a demo. Apache 2.0 open weights.
A generative world model that renders playable first-person environments at 60fps on a consumer RTX 3090 — no game engine.
Key specs
| Parameters | 1.2B parameters |
|---|---|
| Resolution (high tier) | 720p at 60 FPS |
| Resolution (broad) | 360p |
| Gpu target | RTX 3090+ |
| Training data scale | 100x more than v1 |
What is it?
Waypoint-1.5 is Overworld's second-generation real-time video world model, released April 9, 2026. It generates interactive first-person game-like environments entirely through neural rendering — no game engine or rasterization pipeline. Two model tiers: a 720p model targeting RTX 3090–5090 desktop GPUs at up to 60fps (56fps on RTX 5090 reference), and a 360p model targeting gaming laptops and broad consumer hardware. Apache 2.0 license, open weights on HuggingFace.
How does it work?
The 1.2B parameter model was trained on roughly 100x more data than the original Waypoint, enabling significantly more coherent environments with consistent object behavior and motion. Efficient video modeling reduces redundant computation across frames to hit real-time targets. The Biome runtime is the local inference engine; users download a one-click installer, load the checkpoint, and get keyboard/mouse input. A hosted version at overworld.stream requires no setup.
Why does it matter?
Previous world models generated plausible-looking environments but felt like slideshows or blurry demos. Reviewers specifically called Waypoint-1.5 'playable' and compared the shooting mechanics to an actual first-person shooter. Consumer-hardware accessibility makes this the first world model most ML practitioners can run locally — a qualitative threshold that matters for research, not just consumer interest.
Who is it for?
Researchers studying world models, game developers exploring neural rendering, and AI practitioners curious about interactive generation