Google DeepMind releases Genie 4 world model with audio and steering
Google DeepMind has announced Genie 4, a real-time world model that generates interactive 3D worlds from text or image prompts. The model is first available as a research preview for trusted testers. Google says it will roll out to Google AI Ultra subscribers in Project Genie, and to early access partners of the Gemini API and Google AI Studio, in the coming weeks.
Genie 4 builds directly on the Gemini model family. DeepMind says Gemini Omni anchors the simulations in real-world appearance and behavior. Gemini 3.8 Flash works as a guiding agent in the loop, observing the scene, breaking goals into steps, and checking frames to keep the simulation coherent. This integration is the main reason the model can handle real-time steering and longer scenes.
In one line: Google DeepMind released Genie 4, a real-time world model that generates interactive 3D scenes with native audio, more realistic physics, and natural language steering, first as a research preview for trusted testers.
What changed from Genie 3
Most of the updates focus on realism and control. DeepMind lists four capabilities in its announcement: natural language steering, more realistic physics and human motion, native audio, and reasoning with Gemini models.
- Natural language steering lets a user change the weather, move the camera, or add a new person or object while the simulation runs. Gemini adapts the prompt to match the surrounding scene.
- Physics and human motion have improved. DeepMind says Genie 4 simulates aerodynamics, collisions, and humanlike characters that move and respond naturally.
- Native audio is new for Genie. Worlds carry ambient sound such as wind or waves. Characters can speak, and dialogue adapts to user inputs in real time.
- Gemini models handle planning. Gemini 3.8 Flash observes the scene, splits a goal into steps, and reviews incoming frames. If the scene shifts, it adapts the plan.
TechCrunch reports that Genie 4 can keep a scene consistent for up to two minutes, double the limit of its predecessor. The Verge also notes the two minute figure.
How Gemini steers the simulation
The DeepMind blog explains that Gemini 3.8 Flash works as a guiding agent. It continuously observes the scene, breaks goals into a sequence of steps, and checks that each step succeeded. This is not a simple text-to-video pipeline. The model loops between reasoning and rendering.
Google is also testing Genie 4 with Gemini 4 Argon for robotics scenarios. In one example, Argon breaks down a user task into steps. Those steps are interleaved as text captions to Genie 4 while the simulation renders. The benefit is that the simulation can be adapted as the task progresses. This could become a way to test robotic scenarios in imagined worlds before running them in the real world. TechCrunch quotes Tejas Kulkarni, research director at Google, saying the future will increasingly hold hybrid world models where Gemini writes code and text and Genie simulates reality at different levels of abstraction.
What early testers are trying
DeepMind says trusted testers come from creative industries, journalism, education, gaming, and engineering. Their reported use cases include:
- Pre-visualization for filmmakers. Directors can step into a live simulation and run multiple takes, adjusting camera angles, lighting, and dialogue on the fly.
- Turning a single 2D archival photo into a grounded simulation of that place. Photojournalists are testing this to understand historical spaces or visualize events like flash floods.
- Game prototyping. Developers go from an idea to an explorable prototype before writing engine code.
- Simulation for robotics and other enterprises.
TechCrunch adds that Google DeepMind engineer Diego Rivas sees the real time aspect as valuable in ideation and storyboarding. He also cautions that the end-to-end product experience is not clear yet. The Verge and Ars Technica both note that details and object permanence can shift as simulations run longer. Audio accuracy and legible in-world text also remain open problems.
What I would do with it
For our work at Latent, nothing changes immediately. We build RAG chatbots and internal AI rollouts for regulated clients. Genie 4 is a world model, not a retrieval or coding model. It is also not generally available yet, and the known drift in details would be a problem for regulated workflows that need consistent outputs.
What I would watch is the Gemini in the loop pattern. Genie 4 uses Gemini 3.8 Flash to plan, verify frames, and adapt on the fly. That is the same shape as the agentic coding loops we already run for client delivery. If the same steering approach moves into document or process simulation tools, it could be useful. For now, I would join the waitlist only if it does not take time away from client work. Nothing to build yet.
Sources
- Google DeepMind: Genie 4: A new generation of world simulation
- TechCrunch: Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering
- Ars Technica: DeepMind unveils Genie 4 world model promising more realistic physics and humans
- The Verge: Google's next Genie AI model generates more realistic experiences