Google released Genie 4, a world model with real-time audio

A world model that forgets its own scene every thirty seconds is a tech demo. Google DeepMind said on Tuesday that its new one keeps the scene stable for up to two minutes. That is double the last version. The model is called Genie 4. It is in research preview now, for trusted testers.

Genie 4 is Google's fourth world model. The last one, Genie 3, launched in research preview a little over a year ago. The new one adds real-time audio to match the simulated world. The audio can include music, ambient noise, and spoken dialogue. Users can type a prompt mid-simulation. That prompt can change the weather, swap the camera angle, or add a new person or object. These claims are Google's own, from the DeepMind page for the model.

What the research experiment showed

Google ran an experiment that tied Genie 4 to its Gemini 4 Argon model. Argon broke a task into steps. It sent those steps to Genie 4 as text. Genie 4 simulated each step in real time. Google says this is a way to test robot scenarios in simulation first.

Tejas Kulkarni, a research director at Google, described the pattern at a press briefing. He said teams will see more "hybrid world models" soon. In one, Gemini writes code and text. Genie reads that and simulates reality at different levels. That is Kulkarni's own framing, not an outside benchmark.

What Google says it still cannot do

Google was open about what the model still gets wrong. Details and objects can drift when a simulation runs too long. Audio accuracy is still an area of research. In-world text is often not legible. The DeepMind page names each. The model also cannot simulate real-world places with perfect accuracy.

Every clip Genie 4 generates will carry SynthID. That is Google's watermark for AI-generated content. The DeepMind page confirms this.

How to try it, and when

Genie 4 is a research preview for trusted testers today. In the coming weeks, Google AI Ultra subscribers will get access. The entry point is Project Genie. That is Google's tool for making interactive worlds from a text or image prompt. Early access partners of the Gemini API and Google AI Studio will also get access. TechCrunch and the DeepMind page both say so.

Trusted testers told Google they see a near-term use in film prototyping. Diego Rivas, a DeepMind engineer, named ideation and storyboarding as the strongest fits. That is a Google speaker's framing of what trusted testers reported.

A team building with AI should read Genie 4 as a direction, not a product yet. The only commercial surface is Project Genie, inside a paid Google AI Ultra subscription. No API price is published. The research angle is the pair: a world model and a reasoning model, composed. That shape fits the agent evaluation work most teams already do.

Source

Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering, TechCrunch, by Rebecca Bellan. Genie 4, Google DeepMind.