World-model startup General Intuition raises $220M at $6.2B

General Intuition has raised $220 million at a $6.2 billion valuation, the company announced on September 29, a year after it spun out of Medal, the free game-clip platform whose footage it trains on.

The round's investors include Valor Equity Partners, Atreides, Seven Seven Six, Point72, Khosla Ventures and General Catalyst. General Intuition says the money goes to hiring more AI researchers, and it opened a waitlist for a commercial version of the technology alongside the announcement.

The company's argument is about where training data comes from. Models that control robots need footage of the task before they can learn it, and recording that footage by hand is slow. General Intuition builds world models that generate the footage instead, so a robotics team can evaluate a policy inside a simulation rather than on a robot.

Its current model, MIRA, is described by the company as a 5-billion-parameter diffusion transformer paired with a 600-million-parameter video representation codec that simulates four-player Rocket League matches at 20 frames per second. SiliconANGLE reports that, according to General Intuition, it renders 720 by 576 pixels on a single Nvidia B200 GPU.

What makes MIRA interesting to read about is what it does not have. There is no physics engine, no rendering engine and no explicit 3D representation in it: the model learns the game's rules from video and controller inputs. General Intuition says it can run indefinitely without diverging, a stability problem that causal video models usually solve with multi-stage training.

The lab published the model openly in June, releasing a 1,000-hour slice of training data at 720p with paired action streams and physics states, plus training and inference code. The technical report credits researchers at General Intuition, Kyutai and Epic Games, whose game is the test bed.

The limits are stated by the company itself. MIRA is a research demonstration that generates footage of one video game, and it is not being used to develop Rocket League. Videos of factory floors, of a person assembling something, of traffic: those are the domains the technique has to reach before the pitch holds.

For video creators the shift is worth watching even at that distance. The most valuable footage on this beat may soon be the boring kind, and the labs buying it are valuation-heavy and compute-hungry rather than audience-facing. SiliconANGLE notes the company is testing the commercial product with robotics, simulation and entertainment customers.

MIRA's dataset and code are public; the commercial version is in limited testing with a waitlist open.