Google DeepMind has shown the next version of its interactive world model. Genie 4 arrived on October 27 as a limited research preview, and the headline changes are easy to describe: the worlds it generates now make sound, you can change them by typing while you are inside them, and they stay coherent for about two minutes, twice as long as Genie 3 managed.
That sounds like a modest list. It is not, because each item attacks a different reason world models have felt like demos rather than tools.
The short version
- Genie 4 is the first real-time world model to generate native audio, including music, ambient sound and dialogue that reacts to you
- You can steer the world with text mid-simulation: change the weather, move the camera, add people or objects
- Scenes stay consistent for about two minutes, double Genie 3’s limit, though detail can still drift
- Access is limited to trusted testers for now, with a wider rollout reportedly planned for Google AI Ultra subscribers and early API partners
- Google has not published resolution, frame rate or benchmark numbers for Genie 4
What a world model actually is
A normal video model produces a clip and stops. A world model produces a place you can act in. You press a key, the character moves, and the model predicts the next frames on the fly, keeping track of what is where. Genie 3 already did this at 720p and 24 frames per second from a text prompt. The hard part has always been memory: walk away from a tree, turn around, and an older model might have quietly deleted it.
Genie 4 improves that memory and adds two things Genie 3 lacked. One is sound. The other is a way to talk to the simulation while it runs.
Native audio changes how it feels
Earlier world models were silent, which made even good footage feel like a tech demo. Genie 4 generates audio with the same system that generates the picture, so footsteps, wind and background music come from the world itself rather than being pasted on afterward. Characters can also speak, and that dialogue responds to what you do.
This matters for more than atmosphere. If you want to use a simulated world to rehearse a film scene or test a game idea, sound carries a lot of the mood. It is also a harder problem than it looks, since audio has to stay in sync with a scene that nobody scripted ahead of time.
Steering the world with plain text
The most interesting feature is live steering. While a scene is running you can type an instruction such as “make it rain,” shift the camera angle, or drop a new character into the street. The world updates without restarting. Under the hood, Gemini Omni reportedly handles how things look and behave, while other Gemini models work as a reasoning layer that keeps the scene coherent and decides what plausibly happens next.
Where Google sees it being used
Google points to four uses: training robots and AI agents in safe simulated spaces, previewing film scenes, prototyping games and building interactive lessons. The robotics angle is the one with the most weight behind it. A robot cannot practice falling over a thousand times in a real kitchen, but it can in a generated one. Google reportedly showed a demo where Genie 4 simulated multi-step tasks that a Gemini robotics model then worked through.
The games angle is more sensitive. Genie 3 was widely blamed for a wobble in gaming stocks when it appeared, and the industry has not stopped arguing about how far to trust generative tools. We covered the other side of that debate when Epic delayed a shooter to go “AI first”. Genie 4 is nowhere near shipping a finished game, but it makes the prototyping pitch more believable.
| Use case | What Genie 4 offers | Open question |
|---|---|---|
| Robot training | Safe, endless practice environments | Does simulated physics transfer to real hardware? |
| Film previs | Quick scene mockups with sound | Two minutes is short for real direction |
| Game prototyping | Playable ideas from a text prompt | Consistency and control over long sessions |
| Interactive learning | Explorable scenes you can reshape live | Accuracy of what the model invents |
The limits Google is upfront about
Two minutes is better, but it is still two minutes. Google itself warns that visual detail and object consistency can degrade over time, which is the polite way of saying the world can still forget things. There are also no published resolution figures, frame rates, parameter counts or benchmark scores, so any claim that Genie 4 is “more realistic” rests on demo footage and Google’s description. Some outlets also disagree on which Gemini versions are involved, and Google has not settled that publicly.
The wider race is moving quickly. Startups are already raising money on this idea, as we saw when a one-month-old company founded by DeepMind alumni was valued at almost $4 billion on world models alone. Google, meanwhile, is stacking this work on top of its latest frontier model, which we looked at in our piece on Gemini 4 Argon and how hard it is to get access to.
What to watch next
- The rollout. Ultra subscribers and API partners are reportedly next, which is when outside testers can check the claims
- Published specs. Resolution, frame rate and a clear consistency benchmark would turn the demo into something comparable
- Robotics results. A peer-reviewed transfer test from simulation to a real robot would be the strongest proof
- Pricing. Real-time generation is expensive, and the cost per minute will decide who can use it
Genie 4 does not make a world model you can live in. It makes one you can listen to and direct, for two minutes at a time. For a field that spent years producing silent clips, that is a real step, and the next question is whether outside testers see the same thing Google’s demo reel does.
Sources and further reading
- TechCrunch: Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering
- The Next Web: Genie 4 builds AI worlds that talk and change on command
- Unite.AI: Google DeepMind rolls out Genie 4 in research preview
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
