MIRA runs four-player Rocket League at 20 FPS on one B200

A 5-billion-parameter diffusion model that runs four-player Rocket League on a single GPU in real time is the sort of result that sounds like a stunt until you look at what the field has been able to do until now. Most published world models are single-agent, offline, or fall apart after a few seconds of interactive rollout. This one, per the arXiv paper, keeps distributional quality stable to five-minute horizons and, the authors say, continues "for hours with no sign of collapse."

The interesting technical claim is not the size, it is the conditioning. MIRA takes action streams from up to four agents at once, rather than treating the other players as environmental noise. That is the difference between a world model of a game and a world model of one player's slice of a game. The authors frame it as the first multiplayer world model for highly dynamic environments, and the choice of Rocket League, a fast physics sim with lots of body-on-body interaction, is a deliberately hard test case rather than a friendly one.

Why this matters if you are not training world models yourself: interactive simulation is the missing piece for a lot of things people want to build on top of AI right now, from training embodied agents to generating game content to building evaluation harnesses that do not depend on the original engine. A model that can be action-conditioned by several agents at once, at 20 FPS, on one B200, moves that piece closer to being usable infrastructure.

The honest caveat is that 20 FPS on a Nvidia B200 is not a consumer setup, and the 10,000 hours of training data came from bots rather than humans, so the distribution the model has learned to reproduce is a bot distribution. What the paper does not give you is training cost, transfer to less constrained games, or how well the rollouts hold up against human replays instead of bot ones. Those are the questions that decide whether this is a general recipe or a Rocket-League-shaped result.

The forward-looking bit is that the authors are releasing the dataset, the training and inference code, and a live demo, so the multiplayer conditioning idea can be tested by other teams on other environments quickly. That is where you will find out whether MIRA is the start of a category or a very good one-off.