# General Intuition releases MIRA, a playable Rocket League world model > Source: > Published: 2026-07-07 01:42:06+00:00 [General Intuition (@gen_intuition)](https://x.com/gen_intuition/status/2074104524596457706) released MIRA on July 6th, a playable multiplayer world model that simulates a four-player Rocket League-style match in real time rather than running the game through a traditional engine. The release also drew [coverage from The Rundown AI](https://x.com/TheRundownAI/status/2074184559768277398). The model, built by [General Intuition](https://www.generalintuition.com/) and [Kyutai](https://kyutai.org/) in collaboration with [Epic Games](https://www.epicgames.com/site/home), is framed as research, not a replacement for Rocket League. The project's [technical blog](https://mira-wm.com/blog-post/) says MIRA generates a consistent 2v2 match across four player views at 20 frames per second, with a 576p output split across the four players. Under the hood, it pairs a 5B-parameter diffusion transformer with a 600M-parameter video representation codec. That choice of demo matters for Pim de Witte, General Intuition's co-founder and CEO, because the company is trying to prove that games can produce training environments for AI systems that eventually act in physical space. De Witte built General Intuition out of Medal, the gameplay-clips platform, and [RuntimeWire reported in June](https://runtimewire.com/article/general-intuition-320m-series-a-pim-de-witte-medal-gameplay-agents) that he co-founded the AI lab with Eloi Alonso, Adam Jelley and Vincent Micheli. The company's core thesis is that player actions, synchronized with video, give models a cleaner causal signal than video alone. MIRA is a narrower and cleaner version of that bet. General Intuition says the model was trained on roughly 10,000 match-hours of recorded 2v2 Rocket League matches generated by self-play between bots, with no human player gameplay or data used for training. Each match produced four synchronized first-person recordings, one per player, paired with aligned action streams. The team also logged physics state through BakkesMod, but says that state was used for evaluation rather than training. The model saw pixels and actions. The training setup keeps the accomplishment specific. Every car in the dataset was driven by [Nexto](https://github.com/Rolv-Arild/Necto), a public Rocket League bot, across three arenas. That removes human behavior from the dataset and gives the model a cleaner distribution, while limiting behavioral diversity. The blog says the team re-encoded the per-player video to 288p at 20 fps and cut it into four-second chunks for training. General Intuition and Kyutai also released the ingredients around the demo. The [mira-wm/mira GitHub repository](https://github.com/mira-wm/mira) contains the training and inference code under an Apache 2.0 license, and its README describes MIRA as a 5B-parameter latent diffusion model that can run a full 2v2 match inside the model at 20 fps on a single GPU. The [Rocket Science dataset](https://huggingface.co/datasets/kyutai/rocket-science) on Hugging Face includes a 1,000-hour slice of match data with all four views, keyboard actions, game events and per-frame game state. Access requires accepting dataset terms, and the dataset card says the Rocket League gameplay recordings are distributed with Epic Games' permission under CC BY-NC-SA 4.0. The demo shows why world models have become a live investment category. MIRA has no physics engine, rendering engine or explicit 3D representation, according to the project blog. It generates video from actions. In the playable demo, cars move, kick the ball, score goals, run out of boost, trigger demolitions and display event messages. The model can also run in autopilot mode because the researchers used action dropout during training, hiding some player inputs and training the model to predict what would happen without them. The limits are just as important. The blog says replays are one failure case because the model has a roughly four-second context window; after a goal, it can generate a plausible-looking replay that does not match the actual play. Hidden information is another. A single-player model has to infer off-screen players and can forget them. The multiplayer model benefits from seeing all four views and all four action streams tiled together, which makes the demo less occluded than a single embodied agent's view of the world. General Intuition is careful to call MIRA a stepping stone to physical AI, where data is slower, messier and riskier to collect. The claim is not that simulating Rocket League equals understanding the physical world. The stronger claim is that game environments let researchers stress-test action-conditioned video models before pushing similar ideas into robotics, driving or industrial simulation. That strategy is now heavily financed. [RuntimeWire reported on June 25th](https://runtimewire.com/article/general-intuition-320m-series-a-pim-de-witte-medal-gameplay-agents) that General Intuition raised $320 million at a $2.3 billion valuation, bringing total disclosed funding to $454 million after a $134 million launch round in October 2025. Axios also reported the Series A size and valuation, naming Khosla Ventures as lead investor with participation from General Catalyst, Hedosophia, Bezos Expeditions, Innovation Endeavors and Nico Rosberg. General Intuition's own site says it raised the Series A to build models that can perceive, predict and act in virtual and physical environments. The July 6th MIRA release gives those investors a public artifact to point at. It is not a general-purpose robot brain. It is a controlled, multiplayer, action-conditioned simulation of one game, with known constraints and visible artifacts. For General Intuition, that is the point: prove the loop from action to video to consequence in a domain where the ground truth is measurable, then use the lesson to chase environments where mistakes cost real money or real hardware.