Kyutai and Gen Intuition Open Source MIRA, a Model That Simulates Rocket League Without Physics Engine
Kyutai Labs and Gen Intuition released MIRA, a generative model that simulates the multiplayer video game Rocket League without a physics engine. Developed with Epic Games, the model consists of 5 billion parameters and generates gameplay video at 20 frames per second on a single Nvidia B200 chip. Four users played a match within the model during testing, and the developers open-sourced the code alongside a technical report and a dataset derived from 10,000 hours of gameplay.
MIRA learns the game's mechanics by analyzing video footage and controller inputs, bypassing the explicit 3D geometry and rule sets usually required to render a playable experience. According to the technical report, the model maintains coherence during active gameplay but struggles with longer temporal dependencies; the network retains context for approximately four seconds and hallucinates events, such as scoring goals, if replays extend beyond that window.
From the sources (3 posts)
@rronak_Anyone who knows me well knows that I’m obsessed with both Rocket League and World Models. Never thought this day would come so fast!! Great job to the GI team. Now will we get AI GTA 6 before GTA 6?
@kylebrussellRT @gen_intuition: Introducing MIRA. A playable, multiplayer world model. A dream of Rocket League. Trained on 10k hours of data collecte…
@therundownaiFour people just played a full 2v2 Rocket League match inside a neural network. Every frame they saw came out of a 5B-parameter model dreaming the whole game, 20 frames a second, on a single Nvidia B200, with "no physics engine, no renderi