An interactive series on arXiv 2607.05352 · General Intuition × Kyutai × Epic Games
MIRA, explained
MIRA is a 5-billion-parameter world model that is a game of Rocket League. It has no physics or rendering engine: it predicts four players' video, 20 frames per second, from their keyboard presses. These chapters take it apart one component at a time, with demos that actually run in your browser: a real DINOv3 encoder, flow-matching models you train yourself, and learned toy world models you can play inside.
real real model/training running in your browser
toy real learned model in a miniature world
viz exact shapes & mechanics, visualised
paper numbers from the report
The whole machine on one page
Click any block to see its job. Press ▶ one tick to follow a single 100 ms step: one new latent, two new frames for each player.
MIRA's pipeline
shapes from the paperThe one-line summary
Encode each player's view into a small DINO-based latent. Let a transformer, steered by
all four players' keys, generate the next latent with a few flow-matching steps. Decode it into 2 frames per view.
Append it to the context and repeat, 10 times a second, for hours.
The three choices the paper says matter most
① predict in the latent of a frozen pretrained encoder (stability), ② diffusion forcing + few-step distillation (long horizons at speed),
③ condition on all four action streams (less uncertainty, consistent views).