Google DeepMind Unveils Genie 4, a World Model That Generates Sound as Well as Scenery

Google DeepMind has announced Genie 4, its new real-time interactive world model, which generates music, ambient sound and dialogue natively alongside the simulated scene, with Gemini models steering the action in the loop.

Model releasesGoogle DeepMindGemini

Google DeepMind has announced Genie 4, its new world model for real-time interactive simulation. Announced on October 27, 2026, Genie 4 is the successor to Genie 3, which DeepMind introduced in August 2025 as its first real-time interactive general-purpose world model. The headline change this time is sound: Genie 4 is the first real-time world model from DeepMind to generate audio natively, producing music, ambient noise and interactive dialogue at the same time as the world itself.

A world model, in DeepMind usage, is a system that simulates an environment and predicts how it responds over time. Users can step into a generated scene, move through it with keyboard controls, and steer what happens next with natural-language prompts. Genie 4 builds on the Gemini family of models: DeepMind says Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models supply the reasoning needed to keep a simulation coherent and continuous.

The practical pitch is that a user can set an event in motion and watch the world respond, including how it sounds and how the people in it move, react and speak. DeepMind frames the work as a step toward artificial general intelligence, arguing that understanding and simulating the world is a crucial part of building it, and that safe, infinitely variable simulations could serve as training environments for robots and agents as well as a medium for new interactive experiences.

Four capabilities DeepMind highlights

DeepMind organizes the announcement around four advances. The first is real-time steering. Users can change a simulation as it runs, for example by changing the weather, moving the camera, or introducing a new person or object, using natural language. Gemini adapts the prompt to fit the surrounding scene, and between user inputs it monitors and steers the scene so the world keeps progressing in a cohesive way.

The second is realism in physics and human action. DeepMind says Genie 4 sets a new standard for realism in real-time interactive simulation, modeling physical scenarios such as aerodynamics and collisions. The most dramatic leap, in its telling, is how the model simulates people: characters that move naturally, respond organically to their environments, and can be steered interactively in real time.

The third is sound. Every world carries the ambient sound of its own setting, from wind in trees to waves at a beach, all generated by the model rather than added afterwards. The frontier here, DeepMind says, is human speech: characters converse and react to user inputs or events in the world, in real time, with dialogue that adapts on the fly.

The fourth is complex, multi-step events, handled by bringing Gemini models into the loop. Gemini 3.8 Flash acts as a guiding agent while Genie 4 renders the world: it observes the scene, breaks a goal down into a sequence of steps, and reviews incoming frames to track progress and verify that each step succeeded. When the world shifts or the user intervenes, Gemini adapts its plan in the moment. DeepMind says this unlocks worlds with reliable causality, procedural behaviors and multi-step reasoning, and that it has also experimented with combining Gemini 4 Argon and Genie 4, including a demonstration in which Argon breaks down a checkers game for the humanoid robot Apollo from Apptronik into real-time steps that Genie renders.

Availability: research preview first

Genie 4 is not broadly available yet. DeepMind says it has started rolling the model out as a research preview for trusted testers, gathering feedback before a wider release. The next stage, according to the announcement, is Google AI Ultra subscribers in Project Genie, along with early access partners of the Gemini API and Google AI Studio. A waitlist for Genie 4 in Project Genie is open.

That staged rollout mirrors how DeepMind handled Genie 3, which arrived as a limited research preview for a small cohort of academics and creators. The company says it has been working with cohorts of trusted testers across creative industries, journalism, education, gaming and engineering to learn how interactive world models can support real work and where they still fall short.

The tester feedback DeepMind reports points to several early uses. Filmmakers are testing Genie 4 as a rapid draft mode for pre-visualization, stepping into a live simulation of a scene and running multiple takes in real time while adjusting camera angles, lighting, dialogue and action. Photojournalists are exploring how a single archival photograph can become a grounded simulation of an environment in seconds, and how populated, dynamic worlds might help audiences experience historical spaces or visualize short-term environmental events such as a flash flood or wildfire. Game developers are testing it as a way to go from an idea to an explorable prototype in seconds, trying out visual styles and mechanics through natural language before writing engine code. Enterprise partners are evaluating it for rapid scenario generation, identifying gaps that would need to close before it can serve as a simulation tool.

Limitations and safety guardrails

DeepMind is explicit that Genie 4 remains early and experimental. While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time, especially as a simulation runs longer, after multiple complex prompt edits, or while navigating through a scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy and long-term consistency are all listed as active areas of research.

On safety, DeepMind says Genie 4 includes strict guardrails consistent with its policies across other models, and that it ran internal safety evaluations before the preview to prevent the generation of harmful or malicious content. The model is being deployed first to trusted tester partners to gather early feedback before a broader release. All video and audio generated by Genie 4 will carry SynthID, the imperceptible digital watermarking technology DeepMind uses to identify AI-generated content across visual and audio modalities.

Looking ahead, DeepMind points to hybrid world models as a promising direction: systems that combine the logical, structured reasoning of code and language models with the rich sensory representation of pixel-level simulation. Bringing Genie and Gemini closer together is described as an early step in that direction, with Gemini mapping structure and planning the narrative timeline behind the scenes while Genie renders the immediate physical reality, pixel by pixel.

For now, Genie 4 is best read as a research milestone rather than a product launch. The capabilities DeepMind is claiming, native audio, more natural human motion, and Gemini-guided multi-step simulation, are substantial, but the company itself cautions that consistency over long sessions is unresolved. The test will be whether the preview, once it reaches Ultra subscribers and API partners, holds up outside the curated demonstrations.