Google DeepMind announces Genie 4, a real-time world model with native audio

Google DeepMind has announced Genie 4, its new world model. The company says it improves realism and real-time interactive simulation. Genie 4 is described as natively multimodal: music, ambient noises and interactive dialogue are generated at the same time as the world itself. DeepMind calls it its first real-time world model to generate audio natively.

The model builds on the Gemini family. Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a simulation coherent and continuous. The idea, in DeepMind's words, is that you set an event in motion and Genie 4 simulates what happens next: how the world responds, how it sounds, and how the people in it move, react and speak.

DeepMind groups the advances into four capabilities. First, real-time shaping: users navigate with standard keyboard controls and steer the scene with text prompts, for example changing the weather, moving the camera or adding a person or object. Gemini adapts the prompt to the surrounding scene and, between inputs, monitors and steers the world so it keeps progressing cohesively. Second, physics and human action: Genie 4 simulates real-world physics scenarios such as aerodynamics and collisions, and DeepMind calls the simulation of people, who move naturally and respond to their surroundings, the most dramatic leap. Third, sound: each world carries the ambient sound of its setting, and interactive dialogue adapts on the fly, with characters conversing and reacting to the user's inputs or to events. Fourth, complex events with Gemini models: Gemini 3.8 Flash acts as a guiding agent in the loop. As Genie 4 generates the world, Gemini observes the scene, breaks goals into steps, and reviews incoming frames to check that each step succeeded. It adapts the plan if the world shifts or the user intervenes.

DeepMind also says it has been experimenting with Gemini 4 Argon and Genie 4 to connect this research with Gemini Robotics. In the example it describes, Argon breaks a user-instructed task into steps, which are fed to Genie 4 as text captions while the simulation renders, so the simulation can adapt as the task progresses. No results are reported for this experiment.

Trusted testers span creative industries, journalism, education, gaming and engineering. Creatives are testing Genie 4 as a "draft mode" for pre-visualization, where filmmakers step into a live scene and run multiple takes, adjusting camera angles, lighting, dialogue and action. Photojournalists are exploring how a single 2D archival photograph can become a simulation of that environment in seconds, and teams are testing how dynamic worlds could show historical spaces or events such as a flash flood or wildfire. Game developers are going from an idea to an explorable prototype through natural language prompts before writing engine code. Enterprise partners are evaluating it for rapid "zero-to-one" scenario generation and pointing out gaps to fix.

Looking ahead, DeepMind sees hybrid world models as a promising path: they would combine the structured reasoning of code and language models with pixel-level simulation. Gemini would map the structure and plan the narrative timeline, while Genie renders how the world looks, sounds and responds.

On safety, DeepMind says Genie 4 has strict guardrails consistent with its policies for other models, was put through internal safety evaluations before the preview, and went to trusted testers first. All video and audio it generates will carry SynthID watermarks.

The company is open about limits. Genie 4 keeps environments coherent over short sessions, but visual details and object permanence can shift or drift, especially in longer runs, after several complex prompt edits, or when moving through the scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency remain active research areas.

Genie 4 is out as a research preview for trusted testers and is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. A waitlist for Project Genie is open.

Key facts

Genie 4 is Google DeepMind's new world model, which DeepMind calls its first real-time world model to generate audio natively: ambient sound and interactive dialogue.
Users move with keyboard controls and steer scenes with text prompts such as changing weather, camera or adding people; Gemini adapts prompts to the scene.
Gemini 3.8 Flash acts as a guiding agent in the loop, tracking goals and checking frames; Gemini 4 Argon is used in robotics-related experiments.
Availability is a research preview for trusted testers; Google AI Ultra subscribers in Project Genie and early Gemini API and Google AI Studio partners are next, "coming soon".
DeepMind says all output carries SynthID and admits drift in details and object permanence over longer sessions.

Why it matters

DeepMind presents Genie 4 as a step in simulating the world, which it calls a crucial aspect of building AGI. The new elements are native audio, including interactive dialogue, and a Gemini model acting as a planner inside the generation loop, so a simulation can follow multi-step goals rather than just drift. DeepMind also frames the Genie plus Gemini pairing as an early move toward hybrid world models, and says simulated worlds could serve as training environments for robots and agents.

Who it affects

For now, trusted testers in creative industries, journalism, education, gaming and engineering, plus enterprise partners evaluating scenario generation. Next in line are Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio. Filmmakers, game developers and photojournalists are the examples DeepMind gives for pre-visualization, prototyping and turning archival photos into simulations. Robotics and agent researchers are the other audience DeepMind names.

How to use it

There is no general access yet. DeepMind says the model is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio, and it has opened a waitlist for Genie 4 in Project Genie. In use, you navigate with standard keyboard controls and steer the scene with text prompts. No pricing or concrete release date is given.

How solid is it

This is a first-party announcement from Google DeepMind, backed by recordings of real-time interactions in the post. It gives no benchmarks, resolution, frame rate, latency or session-length figures, and no comparison with Genie 3 or other world models. Claims such as "unprecedented realism" and the leap in simulating people are DeepMind's own wording. DeepMind itself calls Genie 4 an early, experimental model. The robotics example with Gemini 4 Argon is described only in prose, with no results.

Risks and caveats

DeepMind lists limits itself: visual details and object permanence can shift or drift over time, especially in longer sessions, after multiple complex prompt edits, or when navigating. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are still active research. On safety, it cites internal evaluations, guardrails consistent with its other models, a staged rollout to trusted testers first, and SynthID watermarking on all video and audio. Tester use cases such as the archival photograph simulation are described as exploration, not delivered results.

“While Genie 4 is a major research milestone, it remains an early, experimental model.”

— Google DeepMind, Genie 4 announcement