Google DeepMind's Genie 4 simulates worlds in real time — with sound and talking characters

Google DeepMind has introduced Genie 4, its first real-time world model that generates sound directly — calculating interactive environments frame by frame and producing matching background audio and character dialogue on the fly. The successor to Genie 3, which DeepMind showed in August 2025, is available now only to select testers as a research preview. It's set to reach Google AI Ultra subscribers in the Project Genie web app soon.

The release matters beyond the demo videos. World models, once they mature, could serve as training environments where robots and AI agents learn from simulated experience rather than just from data — a shift that would touch both DeepMind's robotics program and the broader agent ecosystem being built around Gemini.

What does Genie 4 actually do?

Genie 4 doesn't produce finished videos. It calculates a scene frame by frame, reacting to input as it goes, according to DeepMind. Users move through the scene with the keyboard and steer events with text prompts.

The model marks a first for the company in two ways:

• Sound generated directly. Each world has its own background audio, such as wind in the trees.
• Dialogue in real time. Characters are supposed to talk to each other and react to user input or events as they happen.

DeepMind sees the biggest gain in realism in how the model renders people. Characters are supposed to move naturally and react to their surroundings, and users should be able to control them directly.

That combination — a persistent, explorable world plus audio plus talking characters — is what separates Genie 4 from a text-to-video generator. You don't watch a clip; you act inside a simulation that responds.

The takeaway: Genie 4 turns a generated scene into something closer to a playable environment, complete with its own soundscape and characters that speak.

How do Gemini models keep the world coherent?

Genie 4 builds on capabilities of the video model Gemini Omni so the world looks and behaves realistically, DeepMind says. But rendering is only half the system. Other Gemini models help plan the simulation to keep it consistent.

The mechanism works like this: if a user changes the weather with a text prompt, Gemini adapts the instruction to the running scene. Gemini also steers events between inputs so the world keeps developing in a way that makes sense.

Deepmind explains the process with multistep tasks. Gemini 3.8 Flash watches the scene continuously, breaks a goal into steps, and uses new frames to check whether each step worked. If the world changes or the user steps in, the model adjusts its plan. It's a feedback loop: generate, observe, revise.

The takeaway: a language model acts as the simulation's stage manager, breaking goals into steps and correcting course when the world or the user deviates from the plan.

What's the robotics connection?

DeepMind is also testing Gemini 4 Argon to tie its Genie work to its robotics research, Gemini Robotics. In one demo, Argon breaks the task of playing checkers into steps. Genie 4 receives these as text instructions and uses them to simulate Apptronik's humanoid robot Apollo playing the game — with no real robot involved.

That demo points at the research stakes. A world model that can simulate a specific robot's actions could let engineers test behaviors in software before running them on expensive hardware. It also fits DeepMind's stated long-term direction: the company sees pairing Gemini with Genie as an early step toward hybrid world models that combine the reasoning of language and code models with pixel-level simulation.

The takeaway: simulating Apollo in Genie 4 previews a future where robots rehearse tasks in generated worlds rather than physical ones.

What are the limitations?

DeepMind calls Genie 4 an early, experimental model, and the company is candid about where it breaks down. The world stays consistent over short sessions, but visual details can shift and objects don't always persist in longer simulations or after several complex prompt changes.

DeepMind says it's still researching three open problems:

• Complex physical interactions
• Accurate real-time sound
• Readable text

Those are not edge cases. Object persistence and physical plausibility are exactly what a world model needs if it's going to serve as a training environment for robots or agents — a simulation that forgets its own objects teaches the wrong lessons.

The takeaway: the model holds together in short bursts, but long sessions and layered prompt changes still expose seams.

Who gets access, and how is it watermarked?

For now, Genie 4 is only available to select testers as a research preview. Access is set to expand in stages:

• Google AI Ultra subscribers, via the Project Genie web app, "soon"
• Partners with early access to the Gemini API
• Partners using Google AI Studio

Google says all videos and audio tracks produced by Genie 4 will carry its SynthID watermark, which invisibly marks AI-generated content. That detail matters as generated audio becomes more convincing — watermarking is currently the industry's main tool for distinguishing synthetic media.

The takeaway: access starts narrow and widens gradually, and every output ships with an invisible provenance mark.

Why does a real-time world model matter?

The idea isn't new — agent designs that plan step by step inside simulated environments predate large language models by decades. What's changed is the substrate. Older agent research ran on hand-built simulators. Genie 4 generates the environment itself, on demand, and adapts it while it runs.

DeepMind frames the pairing of Gemini's planning with Genie's pixel-level simulation as the early step toward something bigger: hybrid models that reason in language and code while simulating consequences in generated worlds. If that line of research holds up, training environments stop being a bottleneck that engineers build by hand and become something a model can produce.

For now, the model's own limits — drifting details, vanishing objects, unreliable physics over long sessions — define the gap between the demo and the destination. The direction is clear; the durability isn't yet.