Oasis is the first playable, real-time, open-world AI model jointly launched by Decart and Etched, positioned as an interactive video experience generated entirely by AI. It is officially called "a universe in a Transformer" and is the first step toward research on more complex interactive worlds.
In terms of use, Oasis receives user keyboard input and generates game content in real time, covering physics, rules, and graphics. Users can move, jump, pick up items, break blocks, and more, all without relying on a physics engine, driven only by the foundation model. The model can understand complex internal mechanics such as buildings, lighting physics, inventory management, and object understanding.
Its main capabilities include: placing non-cubic blocks, understanding lighting physics, interacting with animals, eating food to restore health, using a shovel faster than hands, and more. Oasis can also output diverse scenes, locations, and objects. This versatility leads the official team to believe that, with limited additional training, Oasis can be adapted to generate a large number of new maps, experiences, features, and modifications.
In terms of technical architecture, the model consists of two parts: a spatial autoencoder and a latent diffusion backbone, both based on Transformers, with the autoencoder based on ViT and the backbone based on DiT. Unlike action-conditioned world models such as GameNGen and DIAMOND, Oasis chooses Transformers to ensure stable and predictable scaling, and performs fast inference on Etched's Transformer ASIC (Sohu). Unlike bidirectional models such as Sora, Oasis generates frames autoregressively, and each frame can be conditioned on user input, thereby enabling real-time interaction. The model is trained using Diffusion Forcing to denoise with independent per-token noise levels, supporting novel decoding schemes.
To ensure temporal stability over long time horizons, the official team adopted a dynamic noise schedule: adjusting noise according to a plan during inference, injecting noise in the first diffusion forward pass to reduce error accumulation, and then gradually denoising afterward, allowing the model to retain high-frequency details from previous frames and improve consistency.
In terms of performance, Oasis outputs in real time at 20 frames per second. Current state-of-the-art text-to-video models with similar DiT architectures (such as Sora, Mochi-1, and Runway) may take 10 to 20 seconds to generate one second of video even on multiple GPUs; Oasis, by contrast, must generate one frame every 0.04 seconds, more than 100 times faster. With Decart's inference stack, the model reaches real-time frame rates, achieving real-time interactivity for the first time.
HardwareIn terms of hardware, Oasis is optimized for Etched's Transformer ASIC (Sohu). Sohu can scale to next-generation models with 100B+ parameters and supports 4K resolution. Oasis's end-to-end Transformer architecture is extremely efficient on Sohu, and can serve more than 10 times the number of users even on 100B+ parameter models.
The official team also released Oasis's code, a locally runnable 500M parameter model weight, and an online demo of a larger checkpoint.
In future exploration, the official team noted that the model still has limitations: distant video is sometimes blurry, temporal consistency of uncertain objects, domain generalization, precise control of the inventory, and long-context difficulties. The official team hypothesizes that most of these problems can be solved by scaling up the model and dataset, and is currently developing in this direction and optimizing training efficiency.