Genie 4 world model: DeepMind Adds Audio and Live Steering

The Genie 4 world model is Google DeepMind's answer to a simple question: what if you could walk into a video and change it while it plays? Announced on October 27, Genie 4 generates photorealistic, interactive worlds in real time, adds its own audio, and lets you steer the scene with text prompts. It is out now as a research preview for trusted testers, with wider access promised soon.

What actually happened

DeepMind's post, signed by Raia Hadsell, Diego Rivas and Hang Qi, describes Genie 4 as natively multimodal. You can set an event in motion and the model simulates what happens next: how the world responds, how it sounds, and how the people in it move and speak. Users move with standard keyboard controls and type prompts to change the weather, move the camera or add a new person or object mid-scene.

The big architectural change is how closely it works with Gemini. According to Google, Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models handle the reasoning that keeps a scene coherent. TechCrunch describes Gemini as a director that plans steps and checks frames while Genie draws the world. Between user inputs, Gemini also steers the scene to keep it moving in a sensible direction.

DeepMind also showed a robotics experiment. TechCrunch reports that a Gemini 4 Argon model broke a task into steps, and Genie 4 simulated those steps in real time. That is a hint at the long-term goal: testing robot behavior in simulation before it happens in a real room. Google research director Tejas Kulkarni told reporters he expects more "hybrid world models", where Gemini writes code and text and Genie takes that in real time to simulate reality at different levels of detail.

How the Genie 4 world model compares

Genie 3 arrived as a research preview a little over a year ago. Google says Genie 4 improves photorealism, movement and physics and adds native audio. The two-minute consistency window is the most practical upgrade, since earlier world models tended to drift within seconds.

The field is crowded. TechCrunch points to Runway's GWM Worlds 2, a playable model with audio released a couple of months ago, and notes consolidation: World Labs being acquired by AMD, Amazon in talks to buy Decart, and Rivian buying Third Dimension for self-driving simulation. We looked at another contender in Odyssey-3 world model, and at the Gemini model used in the robotics demo in Gemini 4 Argon.

A world model that remembers the room for two minutes is a toy. One that remembers it for an hour is a training ground.

The honest limits

Google was unusually open about gaps. In the press briefing reported by TechCrunch, the team said details and objects can still drift the longer a simulation runs, audio accuracy is not solved, and in-world text is not reliably legible. Diego Rivas stressed it is early and that the final product experience is still unclear.

Testers are pointing to a use Google did not lead with: film prototyping. Rivas said real-time generation is most useful in ideation and storyboarding, where speed keeps creative flow going. That is a more grounded pitch than "step toward AGI", which is how Google frames the bigger picture.

What to do as a reader

If you want to try the Genie 4 world model, join the waitlist through Project Genie. Expect Ultra subscribers and API partners to get it first, with no public date for general access.

Do not expect a game engine. Two minutes of coherence is good for prototypes, mood boards and quick experiments, not for long sessions. Test it on a short scene with a clear goal and judge it on that.

Be wary of crypto tokens or apps claiming Genie 4 access or a Genie token. Google has announced nothing of the kind, and the official path runs only through Google products. Nothing here is investment advice.

Sources: Google DeepMind, Google blog and TechCrunch.