Google DeepMind has released Genie 4, a world model that generates photorealistic, explorable scenes in real time, complete with sound, music and characters who talk back. You can steer it mid-simulation by typing, telling it to change the weather, move the camera or drop a new person into the scene.
DeepMind announced Genie 4 on Tuesday, Oct. 27, posting it on X at 12:10 p.m. ET (9:10 a.m. PT), which was 1:10 a.m. JST on Oct. 28. For now it’s a research preview for trusted testers. DeepMind says it’s coming to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio, and TechCrunch reports that rollout is planned for the coming weeks.
What’s new
A world model is different from a video generator. Instead of rendering a fixed clip, it simulates an environment frame by frame as you move through it, so what you see depends on what you do. Genie 4 adds a few big things over Genie 3, which came out a little over a year ago:
- Native audio. DeepMind says it’s the first of its real-time world models to generate sound itself: ambient noise, music and, most notably, dialogue. Characters can “converse and react” to your input in real time.
- Live text steering. Type a change and Gemini adapts the prompt to fit the scene. Between your inputs, Gemini also watches the world and nudges it so things stay coherent.
- More realism, especially people. DeepMind calls its simulation of human movement “the most dramatic leap,” along with physics like aerodynamics and collisions.
- Longer scenes. TechCrunch reports it can hold a consistent scene for up to two minutes, double Genie 3’s limit.
Under the hood, DeepMind says Genie 4 builds on the Gemini family. Gemini Omni anchors how things look and behave, and Gemini 3.8 Flash acts as a “guiding agent in the loop,” breaking goals into steps, checking incoming frames and adjusting the plan when the user cuts in.
The robotics angle
This is where it gets interesting for anyone who follows Optimus and the humanoid race. DeepMind says understanding and simulating the world is “a crucial aspect of building AGI,” and one of the main uses it lists is creating training environments for robots and agents. In one demo, its Gemini 4 Argon model broke a task down into steps that Genie 4 then simulated in real time: Apptronik’s Apollo humanoid playing a game of checkers. TechCrunch frames it as a way to test robot scenarios in simulation before they happen in the real world.
That’s the same bet every serious robotics and self-driving program is making. Real-world data is slow and expensive to collect. Simulated worlds are cheap and endless, if they’re realistic enough. TechCrunch notes the field is consolidating fast, with Rivian buying simulation company Third Dimension for self-driving testing.
Testers so far include people in film, journalism, education, gaming and engineering. DeepMind says filmmakers are using it as a “draft mode” to run takes and change camera angles on the fly, and photojournalists are testing how a single archival photo can become an explorable scene.
The fine print
DeepMind is upfront that this is early. Details and object permanence can drift the longer a simulation runs or after several complex edits. Physical interactions, audio accuracy, readable text and long-term consistency are all still being worked on. Every video and audio output carries SynthID, Google’s invisible watermark for AI-generated content.
It’s a busy week for AI releases. OpenAI shipped GPT-6.1 Terra for cheaper computer-use agents on Monday. Genie 4 is aimed at a different job: building the worlds those agents, and eventually robots, will learn in.
Sources: Google DeepMind blog, Google blog, Google DeepMind on X, TechCrunch.
— The EVX News team
Featured image: a crowned crane in a Genie 4 simulation, frame from Google DeepMind’s Genie 4 announcement video. Image: Google DeepMind on X.
