Genie 4 World Model Adds Audio, Real-Time Steering, and Two-Minute Scenes

Genie 4 is Google DeepMind's natively multimodal world model for real-time interactive simulation, announced October 27, 2026, that generates synchronized audio, supports natural-language scene steering, and produces characters that move, react to their environments, and speak.

Google DeepMind announced Genie 4 on October 27, 2026, describing it as 'our most realistic and interactive world model yet' in a post on X. The system generates interactive, real-time simulated environments in which audio — ambient sounds, music, and interactive dialogue — is produced simultaneously with the visuals, according to the DeepMind research blog. Google has started distributing Genie 4 to a limited group of trusted testers and plans to widen access in the coming months.

The release is the fourth major iteration of a research line that began with Genie in February 2024, and it arrives at a moment when world models have become one of the most contested frontiers in AI. These systems, which learn the dynamics of an environment and simulate them on demand, are seen by frontier labs as a candidate foundation for embodied AI, robotics training, and the physical reasoning that many researchers believe is required for artificial general intelligence.

Genie 4's entry point into that contest is the Gemini stack. The model incorporates Gemini Omni for photorealism and real-world behavior, with additional Gemini models handling reasoning and scene coherence, giving DeepMind a route from research to product that runs through Project Genie, Google AI Ultra, the Gemini API, and AI Studio.

What is a world model, and why does Genie 4 matter?

A world model is a neural system that learns the dynamics of an environment — how objects move, how light changes, how agents behave — and can simulate those dynamics on demand. Unlike a text-to-video model, which renders a fixed clip, a world model maintains an internal state that persists and evolves, allowing a user to intervene mid-simulation. That distinction is why world models are central to frontier AI research: a credible world model could serve as a training ground for robots and software agents and as a substrate for physical reasoning.

Genie 3, the previous generation, established the core interactivity of the line by letting users steer generated environments. Genie 4 extends that foundation on three axes. It doubles the length of a consistent scene to two minutes, according to TechCrunch's reporting on the launch. It adds native, synchronized audio. And it deepens the behavioral realism of characters, who now move, react to their surroundings, and speak.

The research line has broader context. World models have been a focus across the field since at least the early 2020s, when labs began publishing systems that could predict future frames of video and, later, react to user input. Google's Genie line, first shown in February 2024, brought the concept into the interactive domain, and Genie 3 sharpened the steering capability. Genie 4's additions — audio and dialogue — are the first major multimodal expansion of the series, and they push the model beyond silent video simulation toward something closer to a playable, audible world.

How does Genie 4's real-time natural-language steering work?

Steering is the headline interaction in Genie 4. A user types a plain-language instruction — change the weather, move the camera, introduce a character or an object — and the simulation responds in real time rather than regenerating from scratch. DeepMind describes this as real-time interactive simulation, and the steering is handled inside the generation loop, so the model preserves the state of the world while honoring the new instruction.

The technical challenge is coherence under intervention. Each steering command is a perturbation to a running simulation, and the model must reconcile the new input with everything already established in the scene — the positions of objects, the lighting, the ongoing dialogue. DeepMind says a set of Gemini models supplies the reasoning and scene-coherence layer that keeps these edits consistent, while Gemini Omni contributes the photorealism and real-world behavior that make results visually and physically plausible.

What technical advances separate Genie 4 from Genie 3?

The most visible advance is native audio. Genie 4 generates music, ambient noise, and interactive dialogue simultaneously with the visuals, which means the audio is synchronized to events in the world rather than dubbed on afterward. DeepMind's blog describes the model as natively multimodal, with audio generated as part of the same world-generation process. For filmmaking and gaming, that eliminates a separate sound-design pipeline and makes the simulation feel continuous.

The second advance is behavioral. Characters in Genie 4 move naturally, react to their environments, and engage in dialogue, which requires the model to track social and physical context across frames. This is a meaningful step beyond the minimally animated figures of earlier generations, and it is the capability that matters most for robotics, where agents need to learn in worlds in which other actors behave plausibly.

Genie 3 vs. Genie 4 at a glance

Capability	Genie 3	Genie 4
Native audio	Not generated	Ambient sounds, music, and dialogue generated with visuals (DeepMind blog)
Scene consistency	Up to one minute	Up to two minutes (TechCrunch)
Steering	Basic scene control	Real-time natural-language steering of weather, camera, characters, objects
Character behavior	Limited motion	Natural movement, reactions, and dialogue (Google DeepMind on X)
Access	Research preview	Trusted testers; Google AI Ultra via Project Genie; Gemini API and AI Studio partners (Google)

Which Gemini models power Genie 4, and how do they divide the work?

Genie 4 is not a single model but an orchestrated system. DeepMind says the stack incorporates Gemini Omni for photorealism and real-world behavior, with additional Gemini models handling reasoning and scene coherence. The architecture reflects a broader pattern at Google: rather than routing every task through one massive checkpoint, the company is assembling specialized models into a pipeline, with each component tuned for a distinct responsibility.

Latency is the binding constraint. Because steering happens in real time, the models that interpret user prompts and update the scene must run quickly enough to feel immediate. The Gemini family spans from frontier-scale Omni models to lighter variants such as Gemini 3.8 Flash, and the real-time interaction requirement makes low-latency inference a central design consideration for the system. DeepMind has not published the full routing map for Genie 4, but the division of labor it describes is consistent with the company's broader shift toward model ensembles.

How does Genie 4 change the competitive picture for Google versus Runway?

Runway, the generative video company, is among the most visible rivals in the interactive simulation space, and Genie 4 is Google's direct answer on the creative side of the market. Where Runway has built a standalone platform for filmmakers and studios, Google can attach world models to its Gemini infrastructure and distribute them through Project Genie, the Google AI Ultra subscription tier, the Gemini API, and AI Studio.

That distribution advantage matters because world models are becoming a platform play. A developer who builds a game, a training environment, or a storyboarding tool on top of a world model wants that model accessible in an API they already use, with predictable pricing and a path to scale. Google's cloud and API footprint gives Genie 4 a route into enterprise and startup workflows, and the research-preview rollout is designed to let partners build on the model before it reaches the broader public.

The competitive field extends beyond Runway. A cluster of startups and academic labs are building world models aimed at embodied AI, and the major frontier labs have all invested in simulation research in some form. What distinguishes Genie 4 in that field is the combination of capabilities: real-time interactivity, native audio, character dialogue, and a distribution pipeline that reaches developers through Google's existing AI infrastructure. Rivals may match individual features; matching the full stack is the harder task.

1. Research preview to a limited group of trusted testers, announced October 27, 2026.
2. Google AI Ultra subscribers, who will receive access through Project Genie.
3. Early access partners building with Genie 4 through the Gemini API.
4. Availability in AI Studio for developers and researchers.

What can Genie 4 do for robotics, filmmaking, gaming, and journalism?

DeepMind frames Genie 4 as infrastructure for several industries at once. For robotics, the model offers a safe training environment: robots and software agents can practice in simulated worlds that behave like the physical one, without the cost and risk of real-world trials. The behavioral fidelity contributed by Gemini Omni is what makes those simulations useful for learning, because an agent trained in a world where characters and objects respond plausibly is more likely to transfer its skills to reality.

The agent dimension is explicit in DeepMind's framing. Software agents, like robots, need environments in which to practice — and the interactive, coherent worlds Genie 4 generates are a candidate sandbox for testing agent behavior before deployment. A model that can maintain a scene for two minutes, respond to interventions, and keep characters behaving plausibly is a more demanding training ground than a static dataset, and it is the kind of capability that could shape how the next generation of agents is trained.

For creative industries, the value is speed in the early phases of production. Diego Rivas, an engineer on the DeepMind team, told TechCrunch that the real-time aspect becomes most valuable in keeping a state of flow in the creativity space, citing the ideation and storyboarding phases as the places that benefit most from rapid iteration. Filmmakers can explore a scene from multiple camera angles before committing to a shoot, and game studios can prototype environments interactively rather than rendering static boards.

The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful. For example, in the ideation phase, in the storyboarding phase — anytime you’re kind of in need of rapid ideation.Diego Rivas, Google DeepMind engineer, in TechCrunch

Journalism and manufacturing extend the use cases beyond entertainment. Newsrooms could use interactive simulations to visualize events and explain complex situations to readers, while manufacturers could model factory layouts and production flows before making physical changes. DeepMind's blog describes Genie 4 as advancing safe training environments for robots and agents as well as tools for filmmaking, gaming, journalism, and manufacturing, signaling that the company sees world models as general-purpose infrastructure rather than a single-market product.

How is Google addressing safety and provenance in Genie 4?

Google is releasing Genie 4 through a staged access model. The initial cohort is a limited group of trusted testers, which gives DeepMind a window to observe real-world use before broadening distribution. The subsequent paths — Google AI Ultra via Project Genie and early access partners through the Gemini API and AI Studio — are gated by existing Google accounts and developer agreements, keeping the model inside the company's controlled ecosystem during the preview phase.

For provenance, Google's generative media stack includes SynthID, the company's watermarking system for machine-made content. As interactive world models make synthetic video and audio harder to distinguish from captured footage, watermarking becomes a practical guardrail for journalism, advertising, and platform moderation. Google has not detailed the specific watermarking configuration for Genie 4, but provenance tooling is a recurring component of the company's model releases.

What comes next for Genie and interactive simulation?

The Genie line is evolving toward longer, more coherent, more interactive simulations, and the two-minute ceiling is likely a waypoint rather than a destination. Each generation has extended the duration and fidelity of the worlds the model can hold together, and DeepMind's long-term framing connects world models to AGI research, where the ability to simulate and reason about environments is a core capability. Researchers associated with the project include Raia Hadsell, Diego Rivas, and Hang Qi, and the effort has grown from a research demo into a product pipeline reaching subscribers and developers.

The near-term test is adoption. Whether Genie 4 becomes the default substrate for interactive simulation will depend on how it performs in the hands of trusted testers and early partners now receiving access, and on whether the Gemini-powered architecture can deliver the low latency that real-time steering demands. Google's distribution channels give the model the widest possible runway; the open question is whether its coherence holds up under sustained, unpredictable use.

For the field, Genie 4 is a signal that world models have moved from research curiosities to deployable infrastructure. The combination of native audio, real-time steering, and Gemini-grade photorealism sets a new baseline for interactive simulation and raises the bar for every competitor, from Runway to academic labs publishing their own world-model research. The next 12 months will show how far that baseline moves, and who builds the most valuable applications on top of it.