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)
ReleaseInsight · On RadarThe Decoder→ original

DeepMind Shows Genie 4: A Real-Time World With Sound and Dialogue

Google DeepMind has shown Genie 4, a world model that renders a scene frame by frame as you move through it with a keyboard. For the first time, the model generates its own sound: wind noise and character lines that react to your actions. Gemini 3.8 Flash handles the event planning in the background, splitting goals into steps and checking them against new frames. This is still a closed preview, but such worlds could become training grounds for robots and agents.

AI-processed from The Decoder; edited by Hamidun News
DeepMind Shows Genie 4: A Real-Time World With Sound and Dialogue
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1Access is currently limited to select testers, with Google AI Ultra subscribers to follow
Key metric #2In a demo, an Apptronik Apollo plays checkers in simulation, with no real robot involved
Key metric #3In long sessions, details drift and objects disappear after prompt changes

Google DeepMind has introduced Genie 4, the successor to Genie 3, which the company showed in August 2025. It is a "world model": rather than output a finished video clip, it computes the virtual space while you play. You move through the scene with a keyboard, type text commands, and the world responds. The main novelty of this version is that it is the first DeepMind model to generate sound and dialogue on its own. For game developers, simulator builders, and robotics teams, this signals where interactive environments are heading.

What Genie 4 can do

Think of it less as a video player and more as a living stage set. Every frame is redrawn based on where you turned and what you typed. Each world now has its own ambient audio, for example wind in the trees. Characters can talk to each other and react to your actions or to events in the scene, with lines appearing in real time.

DeepMind is most proud of how the model renders people. According to the company, characters move more naturally, take their surroundings into account, and can be controlled directly. The foundation is the capabilities of the Gemini Omni video model, which handles the plausible look and behavior of the world.

Gemini as the director behind the scenes

The most interesting part is how the "direction" works. Genie 4 itself is responsible for the visuals and sound, but Gemini models help decide what happens next. If you change the weather with a command, Gemini adapts that instruction to the scene already in progress rather than simply switching rain on over everything. Between your actions, it also keeps events moving so the world does not freeze.

DeepMind explains this through multi-step tasks. Gemini 3.8 Flash continuously watches the scene, splits the goal into steps, and uses new frames to check whether each step succeeded. If the world changes or the user intervenes, the model rebuilds the plan. In effect this is the familiar agent loop of observe, plan, act, verify, except the environment is a simulation rather than a browser or terminal.

There is also a link to robotics. DeepMind is testing Gemini 4 Argon to connect Genie with its Gemini Robotics research. In a demo, Argon breaks a checkers game into steps and passes them to Genie 4 as text instructions. The model simulates how Apptronik's humanoid robot Apollo plays checkers. There is no real robot involved.

Limitations and access

The company itself calls Genie 4 an early experimental model. In short sessions the world stays consistent. In long simulations, and after several complex prompt changes, visual details can drift and objects do not always persist. DeepMind is still working on complex physical interaction, accurate real-time audio, and readable text in the frame.

Access is currently restricted: select testers have the preview. The model is promised later to Google AI Ultra subscribers through the Project Genie web app, and to early-access partners of the Gemini API and Google AI Studio. All video and audio will carry the invisible SynthID watermark, which marks AI-generated content.

Why it matters to practitioners

For now Genie 4 is a demonstration, not a product you can build on. But the direction is clear. DeepMind sees the pairing of Gemini and Genie as a first step toward hybrid world models, where the reasoning of language and code models is combined with pixel-level simulation. If such worlds mature, robots and AI agents could be trained on simulated experience, not only on fixed datasets.

For product managers and founders, the takeaway is simple: watch the API and AI Studio. If access opens up, interactive prototypes, training simulators, and test environments for agents will get cheaper. Until then, keep the weak spots in mind: instability in long sessions and physics. Those are exactly the areas where simulation most often diverges from reality.

Story

Key Questions Answered

What Genie 4 can do?

Think of it less as a video player and more as a living stage set. Every frame is redrawn based on where you turned and what you typed. Each world now has its own ambient audio, for example wind in the trees. Characters can talk to each other and react to your actions or to events in the scene, with lines appearing in real time. DeepMind is most proud of how the model renders people. According to the company, characters move more naturally, take their surroundings into account, and can be controlled directly.

What are the key details on Gemini as the director behind the scenes?

The most interesting part is how the "direction" works. Genie 4 itself is responsible for the visuals and sound, but Gemini models help decide what happens next. If you change the weather with a command, Gemini adapts that instruction to the scene already in progress rather than simply switching rain on over everything. Between your actions, it also keeps events moving so the world does not freeze. DeepMind explains this through multi-step tasks.

What are the key details on Limitations and access?

The company itself calls Genie 4 an early experimental model. In short sessions the world stays consistent. In long simulations, and after several complex prompt changes, visual details can drift and objects do not always persist. DeepMind is still working on complex physical interaction, accurate real-time audio, and readable text in the frame. Access is currently restricted: select testers have the preview.

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…