Google DeepMind just unveiled Genie 4, a world model that generates interactive 3D-style environments in real time, now with native audio and spoken dialogue. You type a prompt, move around with standard keyboard controls, and the world keeps responding as you go. For now it is a research preview limited to trusted testers.
The model builds on the Gemini family. DeepMind says one Gemini model anchors the simulation in real-world appearance, while others supply the reasoning needed to keep a scene coherent over time.
What Genie 4 Can Do
Genie 4 is natively multimodal, which means the sound is generated together with the picture instead of being added afterward. Wind in the trees, waves on a beach and music all come out of the same model that draws the scene.
The company highlights three upgrades over earlier versions of Genie.
- More realistic physics, including aerodynamics and collisions.
- People who move naturally and react to their surroundings.
- Interactive dialogue, where characters talk back and adapt to what you do.
You can also change a scene on the fly with plain language. Ask for rain, move the camera, or drop in a new person, and the model adapts the request to fit the world around it.
How Gemini Steers the Simulation
The most interesting part is the pairing with Gemini. DeepMind says Gemini 3.8 Flash acts as a guiding agent inside the loop. While Genie 4 draws the world, Gemini watches the frames, breaks a goal into steps and checks that each step actually worked.
If you step in and change something, Gemini updates its plan on the spot. That is how the team gets more reliable cause and effect, and multi-step behavior, inside a simulated scene.
A Robotics Test Case
DeepMind also tried Gemini 4 Argon with Genie 4. In one demo, Argon split a task into steps for Genie to render, showing Apptronik’s Apollo humanoid robot playing checkers. The goal is a safe place to test how robots and AI agents behave before they face the real world.
| Feature | What DeepMind says |
|---|---|
| Audio | First real-time world model from DeepMind to generate sound natively |
| Dialogue | Characters speak and react to your input as it happens |
| Control | Keyboard navigation plus text prompts to steer the scene |
| Guide model | Gemini 3.8 Flash plans and checks steps in real time |
| Access | Preview for trusted testers first |
Who Is Testing Genie 4 So Far
DeepMind says it has worked with testers in creative industries, journalism, education, gaming and engineering. Filmmakers use it as a quick draft mode, running several takes of a scene while changing the camera, lighting and dialogue live.
Photojournalists are turning a single archival photo into an explorable scene. Game developers are going from an idea to a playable prototype in seconds, and some enterprise partners are using it to generate test scenarios quickly.
Limits and Safety Guardrails
This is still early research, and DeepMind is upfront about it. Scenes can drift over longer sessions, objects may not stay consistent, and readable text, precise spatial accuracy and complex physics are still works in progress.
On safety, the company says Genie 4 went through internal evaluations meant to block harmful content. Every video and audio clip it makes carries a SynthID watermark, an invisible marker that helps identify AI-generated media.
What This Means for You
You cannot try Genie 4 yet, but it shows where AI video is heading. Instead of watching a clip, you will be able to step inside it. Here is what to keep in mind.
- Creators: expect faster pre-visualization tools, where a rough idea becomes a walkable scene.
- Game makers: prototyping may get cheaper long before final art is ready.
- Everyone else: watch for the SynthID label, since realistic simulated worlds with voices will be harder to tell apart from real footage.
If you are curious about the tools available now, you can compare Veo and Seedance or read how AI image generators work.
Worried about fakes? Learn spotting deepfake video in a few minutes, and find more stories are on our AI news page.
Sources
Primary source: Google DeepMind’s Genie 4 announcement. All details above are company claims from that post.
Conclusion
Genie 4 turns AI video into something you can walk through, talk to and steer, with Gemini acting as the planner behind the scenes. It is a preview with clear limits, but it points toward training grounds for robots and new tools for creators, and DeepMind says wider access is coming.
