Google DeepMind released Genie 4 on Tuesday as a research preview, and the advance it is selling is not sharper pixels but simultaneity: the model generates a world, its ambience, its dialogue and the physics that hold it together at the same time, in real time, while a user types instructions into the scene.
That last part is the difference from a video generator. Genie 4 keeps a receive loop open to the user: change the weather, move the camera, drop a new person into a courtyard, and the simulation adapts between keystrokes. Between prompts, a Gemini model watches the frames and steers the world so it keeps progressing on its own.
DeepMind describes the architecture as a division of labour inside the Gemini family. Gemini Omni anchors appearance and behaviour in something resembling the real world; Gemini 3.8 Flash acts as the guiding agent in the loop, breaking goals into steps, reviewing incoming frames and re-planning when the scene moves. In one experiment attached to Gemini Robotics, a Gemini 4 Argon model decomposed a task while Genie 4 rendered Apptronik's Apollo humanoid playing checkers — a hint at rehearsing robot behaviour in generated worlds before hardware does the work.
The numbers DeepMind puts on the model: coherence for up to two minutes, roughly double Genie 3's limit; native audio that includes music, ambient sound and interactive speech; and characters that move and talk in response to events. Every generated video and audio stream carries SynthID, Google's imperceptible watermark.
The limitations are the familiar ones, and DeepMind is unusually direct about them. Details and objects drift as a simulation runs, especially after repeated prompt edits. Legible in-world text, spatial accuracy and audio fidelity remain research problems. The company is rolling Genie 4 out first to trusted testers across film, photojournalism, education and games, then to Google AI Ultra subscribers in Project Genie and early-access partners on the Gemini API and AI Studio.
The context matters: world models are consolidating fast. Runway shipped its second playable world model in the summer, World Labs is being acquired by AMD, Amazon has been reported in talks for Decart, and Rivian bought Third Dimension to bring driving simulation in-house. Genie 4's pitch to that market is not a game-engine replacement but a draft mode — a place to run takes, test camera angles and prototype mechanics before anyone writes engine code.
DeepMind's research directors call the next step hybrid world models: Gemini reasoning about structure and narrative while Genie renders physical reality pixel by pixel. Genie 4 is the first version where that handshake is visible in the product rather than in a paper.
