MIRA runs a playable 2v2 Rocket League match with no game engine
General Intuition, Kyutai, and Epic Games released MIRA in early July, and it's the kind of thing that sounds fake until you play it. MIRA is a 5-billion-parameter world model that simulates a full 2v2 Rocket League match with no game engine underneath. No physics code computes the ball's bounce, no renderer draws the arena. A latent diffusion model takes the button presses of all four players and generates the next frame of every player's view, 20 times a second, on a single GPU. The code is on GitHub, the training data is on Hugging Face, and there's a live demo you can play in your browser at mira-wm.com right now.
The multiplayer part is the news
Playable neural simulations aren't new. GameNGen ran Doom inside a diffusion model back in 2024 and Oasis did the same for Minecraft. But those were single-player. The model answered to one person and treated everything else as scenery. MIRA answers to four people at once. When one car touches the ball, all four screens have to agree on where it goes, and the model has to work out whose input caused which change. It generates four synchronized first-person views of the same match and keeps them coherent enough that an actual 2v2 game works. That's the hard problem here, and as far as we can tell, nobody has shipped an open version of it before.
The training data is a release of its own. The Rocket Science dataset holds roughly 10,000 hours of 2v2 matches, about 29 TB across more than 15,000 games, with time-aligned 720p video, keyboard inputs, game events, and physics state logged at 120 Hz from all four players' perspectives. Every match was played by bots, specifically the publicly available Nexto bot, so there's no human gameplay in it at all. The code ships under Apache 2.0, the dataset under CC BY-NC-SA 4.0, with commercial use of Rocket League content still needing Epic's separate permission. And Epic's name on this is the detail worth pausing on. The company that sells the world's most famous game engine just helped publish a playable game with no engine in it.
What it can't do yet
The authors are refreshingly upfront about the limits. That "single GPU" is an Nvidia B200, a datacenter card, not something in anyone's gaming rig. The output is 720p at 20fps, well below what the real game does on a mid-range PC. The context window is short, so the model can lose track of the score and the clock. The ball sometimes drifts toward a goal on its own, cars occasionally twitch without input, and rare events are undertrained because bots rarely produced them. It simulates one game, learned from one bot's style of play. None of that diminishes the result. It just tells you where on the curve we are.
What it means for game creators
Look at the slope. In 2024 a world model ran Doom for one player. In 2025 it was Minecraft. In mid-2026 it's a four-player team sport with vehicle physics, released as open code with data and a browser demo. Engines aren't in danger this quarter, since a B200 per match against a browser tab is no contest. But two lines are converging: engines are absorbing AI-generated content at runtime, like Meshy's runtime-generated mechanics, while neural simulations are becoming playable, multiplayer, and open. Cinevva sits on the browser-first side of that convergence. Our bet is that game production collapses into a browser, ships instantly, and runs on hardware people already have. When neural simulation gets cheap enough to serve, it will slot into exactly that kind of distribution long before it slots into a 100 GB install. The scarce input doesn't change either way: something worth playing.