Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableTechCrunch→ original

Google DeepMind Releases Genie 4: A World With Sound, On-the-Fly Edits, and Scenes Up to 2 Minutes

Google DeepMind has released Genie 4, a "world model" that builds a playable scene from a text prompt and lets you change that scene on the fly. Mid-simulation, you can type "make it rain" or "add some pedestrians," and the model generates matching music, ambient noise, and dialogue. The scene stays coherent for up to two minutes, twice as long as Genie 3. It is still a preview, but studios and robotics teams should already be looking at how Gemini and Genie fit together.

AI-processed from TechCrunch; edited by Hamidun News
Google DeepMind Releases Genie 4: A World With Sound, On-the-Fly Edits, and Scenes Up to 2 Minutes
Source: TechCrunch. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTechCrunch ↗
Editor-in-Chief
Accessclosed testers now, Google AI Ultra in a few weeks
Key metric #2Gemini acts as the director, planning steps and checking Genie's frames
Weak spotsdetails drift over time, and text inside the world can be unreadable

Imagine launching a game scene from a text description and then, without stopping it, typing "make it rain" or "add some pedestrians." The world changes before your eyes, and music, noises, and character lines play along. Genie 4, the new "world model" from Google DeepMind released as a research preview, can do this. For game developers, simulator builders, and robotics teams, it is a signal that interactive environment generation is moving past the demo stage, and a market is starting to form around it.

What Genie 4 Can Do

A world model is a neural network that renders not a finished video clip but a space that responds to user actions. You move around, change something, and the picture rebuilds immediately. Genie 3 came out a little over a year ago, also as a preview. The fourth version brings several notable improvements.

First, the visuals are the most photorealistic and interactive of any Google world model so far. Second, the model generates sound in real time, adapting it to what is happening: music, ambient noise, and interactive dialogue. Third, users can freely edit the scene during the simulation, changing the weather or camera angle and adding people and objects.

One important numerical parameter stands out. A scene stays coherent for up to two minutes, twice as long as with Genie 3. Google also cites better physics and more "human" characters. The company has not published figures for speed, compute cost, or model size, so the economics of inference cannot be assessed yet.

Gemini as Director, Genie as Cinematographer

For engineers, the most interesting part is the pairing with Gemini. The language model works like a director: it plans the steps and checks the frames while Genie renders the picture. In an experiment connected to Gemini Robotics, the Gemini 4 Argon model broke a task into steps, and Genie 4 simulated them in real time.

The practical implication is that robot routines can be run in a virtual world first and only then deployed in the real one. That is cheaper and safer than breaking hardware. Google DeepMind research director Tejas Kulkarni called this the future of "hybrid world models": Gemini writes code and text, and Genie turns them into a simulation at different levels of abstraction on the fly.

Google names games, interactive learning, and robotics as its commercial applications. For learning, it offers a vivid example: stepping inside a painting with a historical storyline and talking to its figures. Closed testers suggested another niche, film prototyping. DeepMind engineer Diego Rivas explained that real-time mode is valuable where a creative flow must be preserved: at the stage of ideas and storyboards. Later, such sketches could be turned into finished material through Gemini Omni, but that capability is still far off. Google's video model Veo also exists alongside it and is gaining popularity among filmmakers.

Limitations and the Surrounding Market

Google acknowledges the weak spots. The longer a simulation runs, the more its details and objects drift. Audio accuracy is not guaranteed yet, and text inside the world is not always legible. Rivas stressed that this is an early stage and the final user product has not been defined.

Access is also limited. Right now Genie 4 is available to closed testers. Within the next few weeks it will reach Google AI Ultra subscribers through Project Genie, along with early-access partners using the Gemini API and Google AI Studio.

The context matters as much as the model itself. A couple of months ago Runway released its second world model, GWM Worlds 2, which also features sound and responds to user actions. The market is consolidating at the same time: World Labs, led by Fei-Fei Li, is buying AMD, Amazon is negotiating to buy Decart, and Rivian has acquired Third Dimension to build simulations for in-house autonomous vehicle testing.

The takeaway for practitioners: if you build games, educational products, or test robots and self-driving systems, it is worth trying world models now, but not yet putting them into production. Two minutes of stability and detail drift still limit long scenarios. Watch for API access, which will reveal what such simulation costs and whether it is suitable for serious work.

Story

Key Questions Answered

What Genie 4 Can Do?

A world model is a neural network that renders not a finished video clip but a space that responds to user actions. You move around, change something, and the picture rebuilds immediately. Genie 3 came out a little over a year ago, also as a preview. The fourth version brings several notable improvements. First, the visuals are the most photorealistic and interactive of any Google world model so far.

What are the key details on Gemini as Director, Genie as Cinematographer?

For engineers, the most interesting part is the pairing with Gemini. The language model works like a director: it plans the steps and checks the frames while Genie renders the picture. In an experiment connected to Gemini Robotics, the Gemini 4 Argon model broke a task into steps, and Genie 4 simulated them in real time. The practical implication is that robot routines can be run in a virtual world first and only then deployed in the real one. That is cheaper and safer than breaking hardware.

What are the key details on Limitations and the Surrounding Market?

Google acknowledges the weak spots. The longer a simulation runs, the more its details and objects drift. Audio accuracy is not guaranteed yet, and text inside the world is not always legible. Rivas stressed that this is an early stage and the final user product has not been defined. Access is also limited. Right now Genie 4 is available to closed testers.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…