Introducing MIRA. A playable, multiplayer world model. A dream of Rocket League. Trained on 10k hours of data collected with publicly available bots, MIRA learns the dynamics of a four-player game. The model runs in real time at 20 fps, based on the keys you and the other players... press. Built by General Intuition and kyutai, in collaboration with Epic Games. Not used to develop Rocket League. ▶️ Play the demo, read the technical report, and explore the open-source code at At ICML? Find us at Booth 111 to try it yourself and dig into the results with the team.show more

@ricklamers You probably don't want to predict the entire world

Play the four-player demo now.

10k hours sounds like a lot then I remember some people got like 5x that by themselves lmao

Come hang out with us at our ICML booth to play some matches and learn more about MIRA!

i don’t understand what you’re building

Play MIRA at ICML in Seoul with us. Meet the team at booth 111.

Here I was thinking MOBA games would be the first to have a world model

woahh

unless world models lead to net new creativity in gaming or a harry potter off the dome esk design why am i trying to play a knock version of an already existing game

I'm being griefed by my teammate pls add a report button 9/10

If AI progress stalled here I wouldn’t be too upset

wtf

this is incredibly cool.

6 in the morning and I'm playing Rocket League in my browser holy shit this is cool!

this is insanely cool

the most impressive demo so far in 2026 - it's actually playable

insanely cool stuff, as a rl player this is very exciting

i am going to come and destroy MIRA at rocket league 🤣🤣

this is insane 🚀 super cool work

God we are taking off so hard

what the actual

Multiplayer world models before GTA 6.

@0xm1kr @doppelfun

Fascinating

insane

Our friends just played it, absolutely incredible work.

oh.

The name is the same as the AI agents inside Telegram by The Open Platform :)

Curious where medal data was used in this process since it says it was trained off bot data exclusively. “We train on roughly 10,000 match-hours of recorded 2v2 Rocket League matches, generated entirely by self-play between game-playing bots; no human gameplay at all. “

@FelionSpike Wanna try it out? I absolutely called the fact that we would replace entire games with world models with a persistence mechanism

Latency is rough, but impressive work!

watching a model dream the next frame straight from your inputs with no engine underneath is the part that gets me

10k hours just from bot-generated data is wild, curious how it handles real player unpredictability versus what the bots taught it.

This is sooooooo coooool!!!!!

20 fps matters, but recovery from bad player inputs decides if it feels playable