Google Releases Genie 4: A Single Sentence Generates a Virtual World You Can Explore
Google DeepMind has released Genie 4, which generates an explorable virtual world from text input, supports keyboard controls, real-time weather changes and character additions, and adds sound for the first time. The series caused gaming stocks to plunge earlier this year, but Google itself says it is still in the experimental stage.
On Tuesday, Google DeepMind unveiled the next-generation world model, Genie 4, on its official website. Enter a text description and the model generates an explorable virtual world in real time, letting you walk around with the keyboard. While the scene is unfolding, you can also type commands to suddenly change the weather or drop a new character into the scene.
This time it lasts twice as long as the previous generation, with simulated environments running for up to two minutes. It also adds sound for the first time: ambient sounds, music, and dialogue that reacts to user actions.
World models do not generate static videos; they generate environments that can be interacted with in real time. Wall Street has already felt the impact of this kind of technology on the gaming industry. In January this year, Google used Genie 3 to create a prototype product called Project Genie, and when the news came out, gaming stocks fell across the board.
Genie 4 relies on Gemini to maintain coherence in the virtual world. Google said in a blog post that when Genie 4 generates a scene, Gemini observes the incoming frames, understands user goals, breaks them down into steps, and then checks whether each step has been completed. The team experimented with Gemini 4 Argon, but the actual real-time experience is powered by Gemini 3.8 Flash, a lighter-weight workhorse model for everyday tasks.
The new model is currently available as a research preview to a small group of trusted testers. Game developers and filmmakers are already trying to use it to turn their ideas into explorable prototypes. In the coming weeks, Google will open it up to Google AI Ultra subscribers and also provide access to some developers using Google AI Studio and the Gemini API. The company is also researching using it to train robots.
Genie 4 is still an experimental product. Hang Qi, a research engineer at DeepMind, said that running for two minutes does not mean the scene remains perfectly consistent throughout. As simulation runtime extends, visual details and objects may gradually change. Complex physical interactions, clearly readable text, and long-term consistency all still need improvement. Tejas Kulkarni, research director at DeepMind, said that for a virtual world to maintain consistency even for just two minutes, the model must truly understand the physical laws of the everyday world.
All generated audio and video carry SynthID, Google's invisible watermark for AI content identification.
When asked at the press conference about the impact on the gaming industry, DeepMind product manager Diego Rivas responded with restraint: "There are ultimately limits to what you can do in two minutes. We are still in the research and exploration phase and have not truly entered the development phase of a final product."
发布时间: 2026-10-28 07:02