Oasis Decart AI

Oasis reads your key presses and turns them straight into the next frames, with no engine sitting in between. The original release ran at about 20 frames per second at 360p on an Nvidia H100. Decart says that on Etched's Sohu chip, larger checkpoints could reach 4K, which points to where the speed goes next.

The world looks and behaves like Minecraft, so the controls feel familiar. You move, jump, break blocks, open an inventory and use items, and the model reacts to each action. Because it learned game rules, it also handles things like building and lighting, not just movement.

Decart and Etched published the code and the weights of a 500M parameter model, so you can run a smaller Oasis on your own machine. That is rare for a research release of this kind and it is the main reason students and developers can take it apart.

The web demo ships with several starting scenes, from Mountain Meadow to Village Outpost. You can also upload a Minecraft screenshot to seed a new world, which hints at how the model could generate fresh maps with little extra training.

Oasis learned more than pixels. The team reports it picks up physics, especially lighting, along with inventory management and object understanding. That grasp is what separates a world model from a plain video generator, even when the picture is imperfect.

Beyond the hosted demo, you can download the model and run it offline. That gives you full control over the session length and lets you test changes to prompts, seeds and frame handling without a queue.

The project includes a technical report that explains the architecture. Oasis pairs a ViT-based spatial autoencoder with a DiT-based latent diffusion backbone, and it generates frames autoregressively with a method the team calls dynamic noising to fight error build-up.