MIRA
MIRA is a joint General Intuition and Kyutai world model developed with Epic Games. It generates four synchronized views of a Rocket League match from player actions rather than using a physics or rendering engine.
Research release
The training footage comes from bots
Training uses roughly 10,000 match-hours of bot self-play across three arenas. Human player data was not used. The narrow game setting limits what the result establishes about real-world simulation.MIRA: world model, data and limits
The demo and code have different roles
The public demo provides interactive access. The Apache 2.0 repository releases training and inference code; it does not identify a pretrained world-model download in its README. Do not treat the code license as proof that model weights are available.MIRA: official training and inference code
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Provider product |
| Model identifiers | Multiplayer Interactive World Models with Representation Autoencoders |
| Architecture | 5B diffusion transformer and a roughly 600M video representation codec.MIRA: world model, data and limits |
| Demonstrated setting | Four-player 2v2 matches at 20 frames per second.MIRA: world model, data and limits |