MIRA turns video generation into a playable 4-player game

Imagine jumping into a 2v2 Rocket League match with three friends. Cars drift, crash into each other, flip through the air to kick the ball, and custom banners flash "Shot on goal" across the screen. But beneath the surface, there is no Unreal Engine, no physics solver, and no 3D asset pipeline.

The entire server is just an AI generating video frames on the fly. Every time you tap a key, the model calculates what the next frame should look like and renders it in real time.

That is MIRA, an interactive multiplayer world model recently released by researchers. Powered by a 5B-parameter diffusion transformer paired with a 600M video representation codec, the system runs on a single NVIDIA B200 GPU, delivering synchronized 20 fps gameplay split across four 576p player feeds.

Learning physics purely from pixels

The magic of MIRA lies in the fact that it has zero explicit code for physics or rendering. Yet, it respects collision dynamics, tracks nitro boost meters so you cannot accelerate when empty, and faithfully handles car demolitions.

To pull this off, the team trained the system on roughly 10,000 match-hours of 2v2 gameplay. Crucially, none of this data came from human players; it was gathered entirely from self-play matches between advanced Nexto bots. The model was fed only synchronized first-person gameplay video and aligned action inputs, while internal physics states were kept strictly for evaluation.

Long-term rollouts usually kill causal video models, which tend to melt into visual noise after a few seconds. MIRA manages to run indefinitely without diverging, even with a relatively simple recipe. Thanks to action dropout during training, the demo also includes an autopilot feature where the model predicts optimal bot inputs, letting four AI drivers duke it out entirely within the hallucinated video space.

Why it matters

We think MIRA represents a significant engineering proof of concept for physical AI, far beyond its novelty as a web game.

Simulating an existing video game might sound redundant at first—the researchers themselves rightly note that if you want entertainment, you should just play the original game. But video games provide clean, rich sandbox environments to crack the hardest problems in embodied AI and autonomous driving. Testing safety-critical agents in the real world is dangerously expensive, while training inside an accurate world model is both fast and completely safe.

MIRA demonstrates that a relatively compact 5B diffusion model can maintain consistent multi-agent interactions and physical intuition under tight compute budgets. While the current build is constrained to three arenas and a shared 576p resolution, open-sourcing the training code, inference pipeline, and dataset lowers the barrier for the broader AI community to take interactive simulation toward real-world robotics.