Genie 4: DeepMind's world model now generates worlds with sound and dialogue

What it's about

On October 27, 2026, Google DeepMind unveiled Genie 4, its new world model. A world model is an AI system that simulates an environment. You move around in it like in a video game, but every frame is generated by the AI instead of coming from pre-built graphics. Genie 4 is meant to be far more photorealistic than its predecessor and is "natively multimodal". That means music, ambient sounds and dialogue are created at the same time as the world itself.

The model is launching as a research preview for selected "Trusted Testers". In the coming weeks, Google AI Ultra subscribers should get access in "Project Genie", along with early-access partners of the Gemini API and Google AI Studio. Project Genie has a waitlist. Weights will not be released.

What Genie 4 can do

DeepMind names four core capabilities. The demos shown are recordings of real-time interaction, in which users navigate with the keyboard, control the camera and steer the scene with text.

• Real-time control: Use natural language to change the weather or camera position, or to add people and objects. Gemini adapts your prompt to the surrounding scene. Between your prompts, Gemini watches the scene and steers it so the world keeps running coherently.
• Realism in physics and people: Genie 4 simulates physical situations such as aerodynamics and collisions. DeepMind sees the biggest leap in people: characters move naturally, react to their surroundings and can be directed in real time.
• Sound: It is DeepMind's first real-time world model with native audio generation. Every world has its own soundscape, such as wind in the trees or waves on a beach. There is also dialogue that adapts on the fly: characters talk to each other and react to your input or to events.
• Complex events with Gemini: Gemini 3.8 Flash works as a steering agent "in the loop". It watches the scene, breaks goals into sub-steps and checks incoming frames to see whether each step succeeded. If the situation changes or the user intervenes, it adjusts the plan. This is meant to make causality, procedural behavior and multi-step reasoning more reliable.

Genie 4 also draws on capabilities from the Gemini family: Gemini Omni grounds the simulation in real-world appearance and behavior, while other Gemini models provide coherence and continuity.

Robotics experiment

DeepMind tested Genie 4 together with Gemini 4 Argon and Gemini Robotics. Argon breaks a given task into steps, which are fed into the running simulation as text captions. According to Unite.AI, the systems used this to simulate Apptronik's Apollo humanoid playing checkers. The idea is to play through robot scenarios in simulation before they happen in reality. This is research, not a product.

Improvements over Genie 3

Genie 3 was unveiled on August 5, 2025. It ran at 24 frames per second in 720p and stayed consistent for a few minutes, as a limited preview for a small number of researchers and creators. According to TechCrunch, Genie 4 keeps a scene consistent for up to two minutes, twice the predecessor's limit. This apparently refers to scene consistency during free-form edits; the sources give no further details. Audio, free-form editing and more human-like characters are also new. The sources give no figures for Genie 4's resolution or frame rate.

What testers are doing with it

DeepMind is working with testers from the creative industries, journalism, education, games and engineering:

• Film: Genie 4 as a fast "draft mode" for previsualization. You step into a scene and play through several "takes" with different camera angles, lighting, dialogue and action.
• Historical scenarios: A single archived 2D photo is said to become a walkable environment in seconds. Events such as flash floods or wildfires could also be visualized.
• Game prototypes: From idea to explorable prototype in seconds, before a single line of engine code is written.
• Simulation: Enterprise partners are evaluating fast "zero-to-one" scenarios.

In the briefing, DeepMind engineer Diego Rivas said, according to TechCrunch, that the real-time aspect is valuable where creative flow matters, such as ideation and storyboarding. Gemini Omni might later turn such prototypes into finished assets, but that is still a long way off. As a speculative learning example, TechCrunch mentions stepping into a painting and talking to the figures in it.

Hybrid world models

DeepMind names hybrid world models as a future direction: language and code models provide structured logic, while Genie provides the sensory pixel simulation. Gemini plans the structure and the course of the action, and Genie generates image, sound and reaction. According to TechCrunch, Tejas Kulkarni said we will see more and more of these hybrids, in which Gemini writes code and text and Genie simulates it in real time at different levels of abstraction.

Safety and limits

DeepMind cites strict safeguards and internal safety reviews against harmful content, and says the launch with testers is deliberate. All video and audio contain SynthID, an invisible digital watermark for detecting AI-generated content.

Google openly admits the limits: details and object permanence (things staying where they are) can drift, especially over longer sessions, with many complex changes or when moving through the scene. Complex physics, audio accuracy, legible text, spatial accuracy and long-term consistency are open research questions. Rivas stressed that it's unclear what the product experience will ultimately look like.

Our take

Genie 4 shows where world models are heading: away from silent video and toward interactive scenes with sound, speech and planning. TechCrunch places it in a tough field: Runway released GWM Worlds 2 with audio a few months ago, and the market is consolidating. World Labs is reportedly being bought by AMD, Amazon is in talks with Decart, and Rivian acquired Third Dimension. The idea is attractive for robot training and game prototyping, but every claim comes from Google itself, and there are no independent benchmarks. Given the two-minute limit, the drift and access through the pricey Ultra subscription, it's more of a preview than a tool for most users for now.