What Genie 4 Is For

What DeepMind shipped on October 27

Google DeepMind announced Genie 4 as a world model that generates an interactive scene and the sound that goes with it, including speech. You move and aim with ordinary keyboard controls. You change the scene with text while it is running. DeepMind says Gemini Omni anchors how the world looks and behaves, and that other Gemini models help keep the simulation coherent.

This is a research preview for trusted testers. DeepMind says a wider release comes later, starting with Google AI Ultra subscribers in Project Genie and early partners on the Gemini API and Google AI Studio. There is a waitlist. The post does not say Genie 4 is open in ordinary Gemini chat today.

Genie 4 is not Luma Genie. That listing is a different product, for 3D assets. The October 27 model is DeepMind's world simulator.

These are DeepMind's descriptions of its own preview. They are not a CuratedAI test of the clips, the physics, or the waitlist.

Read the steering claims as the lab's own account

DeepMind's case is that a running world can be redirected in natural language: weather, camera, a new person, a new object. The post says Gemini adapts that prompt to the scene already on screen, and that between your inputs it keeps the world moving. The same post says Genie 4 is their first real-time world model with native audio, and that characters can speak and react while the scene runs.

A second loop uses Gemini 3.8 Flash as a guide inside the simulation. DeepMind says that model watches the frames, splits a goal into steps, and checks whether each step happened. A separate experiment pairs Gemini 4 Argon with Genie 4 on a robotics task: Argon breaks a checkers game for Apptronik's Apollo humanoid into steps, and Genie renders them. There is no Gemini 3.8 Flash listing in this directory. Treat both loops as DeepMind's research description, not as a controller you can download.

Every video and audio clip from Genie 4 carries SynthID, DeepMind's watermark. That is a label on the output. It is not a promise that a long session stays consistent.

Match the job to a preview, not to a finished tool

The reason to try Genie 4 is a short scene you want to walk through and redirect. The reason to stay on a tool you already have is anything that has to leave as a file this week.

Use the limits DeepMind already printed

• Keep the session short. DeepMind says a scene can hold together briefly, and that details and object permanence can drift as it runs longer, after several prompt edits, or while you move through the space.
• Do not ask it for readable text. The limitations section names legible text as ongoing research. A sign, a UI, or a title card is the wrong test.
• Do not grade physics from the demo reel. DeepMind says the model simulates collisions and aerodynamics, and also says complex physical interactions are still being researched. Those two sentences are both theirs. A watermelon clip is not a measurement.
• Check sound on a real speaker. Native audio, including speech, is the new claim. Real-time audio accuracy is also on the limitations list. Listen before you plan a dialogue scene around it.
• Confirm you are in the preview. Trusted testers have it. Ultra subscribers and API partners are the next groups DeepMind named. A colleague's screen recording is not access.

Where this release is easy to overread

"A new generation of world simulation" describes a research preview with a waitlist. It does not say Genie 4 replaced Veo, a game engine, or a robotics stack. The tester section is explicit about the current use: draft a scene, walk through a historical photograph, try an interactive idea, sketch a scenario. Each of those is a study, and DeepMind says the model is early.

The same day's other model release is Liquid AI's LFM3, six open-weight models around 2 billion parameters, which you can download now. That is a different job from a world you steer in a preview. Both are in the October 27 briefing.

What to do this week

If you need a short scene you can redirect while it plays, join the Project Genie waitlist and wait for access. Until a session of yours holds the objects and the words you care about, keep publishable footage on the tool you already export from. For the rest of the day, read the October 27 briefing. For the Gemini models already listed here, see Gemini Omni and Gemini 4 Argon. Reviewed October 28, 2026, from DeepMind's Genie 4 post.