Google DeepMind announced Genie 4 on 27 October 2026, opening a research preview of a world model that generates interactive scenes and their sound together. Trusted testers can move through a simulation and alter it while it runs, rather than choosing only what goes into the initial prompt. The company says it is gathering their feedback before offering access through Project Genie, the Gemini API and Google AI Studio.
Key points
- Genie 4 generates visuals and native audio, including ambient sound and interactive dialogue.
- Its simulations can run for up to two minutes, twice Genie 3’s limit.
- Google DeepMind has begun a research preview with trusted testers.
Gemini steers Genie 4 as scenes unfold
A world model makes the next view of a virtual place in response to what happens inside it and what a user does. Genie 4 generates that view frame by frame in real time, without a finished environment waiting behind the screen. Google DeepMind’s recordings show people navigating with a keyboard and using text prompts to change a scene already in progress. Its examples include changing the weather, moving the camera and introducing another person or object.
Google DeepMind says Genie 4 draws on its Gemini models in different ways. Gemini Omni supplies a basis for how things look and behave in the real world, while other Gemini models help the simulation retain its direction as new events occur. The company says Gemini also adapts a user’s instruction to the surrounding scene and steers what happens between instructions. In the recordings, that process accompanies the continuous generation of pictures and sound.
Sketching a scene for a film could become an exercise in moving through it, changing the weather or camera position as it runs, rather than viewing a fixed sequence. Sound and dialogue generated alongside the images would let that exploration include what might be heard. That use follows from the controls and simultaneous audio generation Google DeepMind describes, rather than from a finished film-making product.
Google DeepMind describes the audio as native to the model: it generates ambient noise, music-like sound and interactive dialogue alongside the visual scene. That matters for an environment intended to react to someone moving through it. A change in the scene can call for a change in what is heard, as well as what appears on screen. The company says all the accompanying sound in its published examples was generated by Genie 4.
Genie 4 reaches a two-minute limit
Genie 4 simulations can last up to two minutes, double Genie 3’s limit, Quartz reported. Duration matters because an interactive world must account for where things were and what happened to them as the view changes. Quartz also reported a limit within that running time: objects and details may drift or distort in longer scenes, and reaching two minutes does not guarantee that the setting remains consistent throughout.
Gemini 3.8 Flash acts as a guiding agent that breaks a goal into parts and assesses each generated frame, Ars Technica reported. The arrangement gives the simulation continuing direction as a user moves or revises a prompt. Google DeepMind says worlds still lose track of their starting conditions after a couple of minutes, according to Ars Technica. A guiding model can therefore help maintain continuity without making it indefinite.
Google DeepMind also says Genie 4 handles physical interactions and human movement more convincingly, Ars Technica reported. The examples published with the announcement are recordings of real-time interactions, rather than a reported measurement of physical accuracy. According to Quartz, Google DeepMind’s researchers have yet to establish a way of guaranteeing that generated sound is faithful and writing within simulated scenes is readable. Those limits bear directly on the training environments for robots and agents that Google DeepMind identifies as a possible use: a simulation used for practice would need to behave reliably when an action depends on a detail.
Project Genie starts with trusted testers
Google DeepMind has begun giving Genie 4 to a limited group of trusted testers as a research preview. The company says it is using their feedback before making the model available to consumers and developers, beginning with Google AI Ultra subscribers in Project Genie and early-access partners using the Gemini API and Google AI Studio. Project Genie gives users a way to explore generated worlds, while the API and Studio are the planned routes for partners building with the model.
The company presents world simulation as part of its work towards artificial general intelligence and proposes training settings for robots and agents, alongside new interactive experiences. For developers, Genie 4’s combination of live control, sound and longer scenes expands what can be tried in a prototype. Its preview status and the reported continuity and accuracy limits remain part of that proposition. Google DeepMind has opened a waitlist for Genie 4 in Project Genie.
