AI research group General Intuition, in collaboration with French AI lab Kyutai and Epic Games, has released MIRA, a multi-player interactive world model. The generative game simulator can render 2v2 matches of Rocket League in real time — no physics engine, no rendering engine, no explicit 3D representations. All it needs are the historical frames and player inputs.
https://twitter.com/gen_intuition/status/2074104524596457706
Unlike companies like Odyssey that decouple logic and rendering, MIRA takes a generative simulation approach based on video latent space. With 5 billion parameters, its core design builds the latent prediction space on top of a frozen universal visual encoder: DINOv3-L. By leveraging pretrained visual features, the generated latent states remain more stable within the valid representation space, significantly mitigating visual drift and divergence during long-term predictions.
For multi-screen alignment, MIRA stitches the latent frames from four player perspectives into a unified grid. This lets spatial attention mechanisms operate naturally across viewpoints, improving spatial consistency for vehicles, the ball, and key events. Training with action dropout also helps the system fill in the behavior of vehicles not controlled by input commands when some action streams are missing.
MIRA runs at 20 frames per second on a single Nvidia B200 GPU. The team has open-sourced both training and inference code, and released the Rocket Science dataset — 1,000 hours of matches, totaling about 4,000 hours of video across four perspectives, along with action streams and physical state data. The full model was trained on roughly 10,000 hours of clean gameplay data.