Oasis AI Game

Oasis is a generative world model developed by Decart and Etched that produces interactive, playable game environments in real time by predicting the next visual frame after each player input rather than executing pre-written game logic or rendering from static assets. The architecture uses a Vision Transformer to encode the current game state, a diffusion model to generate the next frame based on that state and the player's input, and a dynamic noising technique to maintain frame-to-frame visual stability. The result is a playable Minecraft-inspired environment where players move with WASD keys and mouse, jump, break blocks, pick up items, and interact with basic physics—all generated by the AI model on the fly at approximately 20 FPS. A Custom Worlds feature, released after the initial launch, allows users to upload their own starting image and have Oasis generate a playable game world derived from that image. The open-source release includes the code and weights for a 500-million-parameter model runnable locally, alongside a hosted demo of a larger checkpoint. Decart's inference engine makes real-time generation possible on NVIDIA H100 GPUs, and the model is architecturally optimized for Etched's Sohu Transformer ASIC, which when released is designed to support 100B+ parameter models at 4K resolution. The project received $21 million in seed funding from Sequoia Capital and Oren Zeev. Oasis is primarily used as a research demonstration and proof-of-concept by AI researchers studying world models, generative video systems, and real-time neural rendering. Developers exploring AI-generated game environments reference the open-source codebase to understand the architecture and experiment with adaptations. Gaming researchers study the Custom Worlds feature as an early indication of how image-to-playable-world generation might develop. Educators and technical communicators use the publicly accessible demo to illustrate the current state of AI world model research to non-specialist audiences. The Wikipedia entry on Oasis notes that Decart did not receive permission from Microsoft—which owns Minecraft—to train on Minecraft footage, raising unresolved intellectual property questions about the training data. Oasis's core technical contribution is achieving real-time interactive generation at 20 FPS—approximately 200 times faster than other diffusion-based video models at the time of release, which generated roughly 0.1 frames per second. This speed differential is what makes the system interactively playable rather than a passive video generation tool. Current limitations include low resolution (360p in the initial demo), visual blurring at distance, object appearance inconsistency across frames, memory loss in longer sessions where the world layout degrades or changes unexpectedly, and a lack of audio. The world model does not maintain a persistent spatial map, meaning turning around can result in a visually different landscape than the one just left. These constraints are consistent with the development team's characterization of Oasis as a proof-of-concept and research platform rather than a consumer game release.