Google DeepMind announces Genie 4 world simulation model

By Raia Hadsell, Diego Rivas and Hang Qi

Google DeepMind has announced Genie 4, a new world model designed to deliver more realistic, real-time interactive simulations. The natively multimodal system can generate music-like audio, ambient sound and interactive dialogue alongside a simulated world.

Genie 4 incorporates capabilities from the Gemini family. Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models provide the reasoning needed to maintain coherence and continuity. Users can set an event in motion and Genie 4 simulates what happens next, including how the environment responds, sounds and how people move, react and speak.

Google DeepMind said understanding and simulating the world is an important aspect of building AGI. It said Genie 4 could support training environments for robots and agents as well as the development of interactive experiences. In that sense, world models are positioned as a counterpart to language models — one modeling how environments unfold in space and time, the other how ideas and instructions are expressed — a pairing DeepMind is explicitly building toward with the Gemini–Genie combination detailed below.

The company has begun rolling out Genie 4 as a research preview to trusted testers while it gathers feedback ahead of a broader release to consumers and developers. The next phase will begin with Google AI Ultra subscribers in Project Genie and early-access partners using the Gemini API and Google AI Studio. A waitlist is available through Project Genie. The staged rollout also gives outside observers a structured way to track progress on the model’s known weak spots, such as long-session consistency and real-time audio accuracy, before general availability.

Genie 4 is designed to provide enhanced photorealism, fluid movement and native audio. Google DeepMind said the accompanying sounds in its demonstrations were produced directly by the model.

Four capabilities highlighted by Google DeepMind

The demonstrations described below show real-time interactions in which users navigate and control the camera with standard keyboard inputs and guide scenes with text prompts.

Real-time world simulation

Genie 4 allows users to modify a simulation with natural-language instructions, such as changing the weather, moving the camera or adding a person or object. Gemini adapts the prompt to the surrounding scene, while also monitoring and steering the environment between user inputs so that the simulation continues coherently. That continuous, steerable quality is what distinguishes interactive world models from pre-rendered video: the environment keeps evolving and responding even between user inputs.

One demonstration uses a reference image as the exact first frame for a scene featuring a floating rectangular platform, weathered concrete edges, wild grasses, ferns, a twisted small tree, a glowing amber sphere and a middle-aged man in a loose brown suit playing a brass saxophone. The requested visual style is a highly naturalistic, low-contrast documentary scene photographed on 16mm film, with subtle grain, muted natural colors, lifted shadows, gentle highlights and diffuse overcast light. The platform floats on a quiet lake surrounded by misty forests, while fog drifts above the water and the amber sphere reflects on the lake.

The saxophone player remains in view as the camera moves independently over the water. The prompt calls for realistic breathing, finger movements, gentle body swaying, spatial continuity, lake ambience, distant wind, soft water movement and realistic saxophone audio, with no background music beyond instruments physically present in the scene.

Other examples include a sunlit hilltop meadow filled with tall green grass, purple, white and yellow wildflowers, low shrubs and distant blue-green hills. A continuous stream of glossy bubblegum-pink satin winds through the meadow in sweeping S-curves, billowing above the flowers like a river of silk.

A separate prompt depicts an aerial view of Antarctica with snow-covered glaciers, jagged mountains and blue icebergs under polar sunlight. Hot-air balloons appear at different distances, including a large cream, red and blue striped balloon. The first-person view is from inside the woven wicker basket of a balloon, looking over its leather-trimmed edge, black padded support poles, metal rigging chains and the dark base of the canopy.

Physics and human action

Google DeepMind said Genie 4 is built on an advanced physical understanding of the world and can simulate scenarios involving aerodynamics, collisions and other real-world physics. The company described its treatment of people as a major advance, saying characters can move naturally, respond to their surroundings and be steered interactively in real time. Physical plausibility is a central benchmark for world models because training robots and agents in simulation depends on simulated actions producing realistic consequences.

Demonstrations include a glossy green-striped watermelon on a weathered stone plinth in a sunlit courtyard, with airborne droplets backlit in slow motion; a maple seed falling from a tree into grass; and a crowded rural Japanese train station in the 1980s. In the station scene, a railway inspector in a dark navy uniform observes passengers waiting, walking, boarding and carrying luggage. A long, deep-red local train occupies the nearest track, while green mountains, cedar forests and humid summer haze surround the station. The prompt specifies an authentic 1980s Japanese atmosphere, 35mm documentary photography, natural lighting, subtle film grain and a dense crowd rather than an empty or sparsely populated platform.

Native audio and interactive dialogue

Genie 4 is Google DeepMind’s first real-time world model designed to generate audio natively. Its environments can include setting-specific sounds, such as wind in trees and waves at a beach. Native generation of sound alongside visuals is a defining trait of real-time world models, removing the need to pair generated footage with separately produced audio. The company also highlighted interactive dialogue, with characters able to converse and respond to user inputs or events in real time.

One demonstration asks Genie 4 to create a fully three-dimensional stained-glass landscape that can be explored as if the camera were moving through a giant crystal window. The river, grass, flowers, trees, mountains, clouds, sky and sun are all made from translucent colored glass, dark lead outlines and layered panels. The prompt requires the world to remain entirely within the stained-glass style, without natural textures, modern objects or ordinary real-world surfaces.

Another demonstration presents a first-person view from a knight riding a gigantic black dragon at 140 km/h through a mountainous racing course. The rider’s face, body and silhouette remain invisible; only the rider’s hands, the dragon’s neck and upper back, the landscape, other dragons and riders, and a castle with catapults are visible. The prompt calls for realistic wingbeats, wind, motion blur and head movement, with audio consisting of wingbeats, high-altitude wind, distant volcanic rumble and crackling embers.

A further example depicts two Korean lovers in their 20s riding a motorcycle across a bridge near Hong Kong at dawn in a 1995 independent romantic drama. The male driver is not wearing a helmet, while the female passenger wears an open-face helmet. The scene is framed as an intimate, melancholic 35mm film shot with black bars, shallow focus, orange streetlights and a dark green sky.

Complex events with Gemini models

Google DeepMind said combining Gemini and Genie 4 adds reasoning for complex, multistep events and allows AI systems to be tested in simulated environments before being used in the real world.

In the real-time setup, Gemini 3.8 Flash acts as a guiding agent. As Genie 4 generates a world, Gemini observes the scene, divides goals into sequences of steps, reviews incoming frames and checks whether each step succeeded. It can adapt its plan when the world changes or the user intervenes, supporting causality, procedural behavior and multistep reasoning.

Examples include a Universal Robots UR5 holding a pipette and dispensing liquid into a laboratory plate, triggering tiny electric sparks. Another prompt describes a first-person journey through an opulent, claustrophobic antique salon. The viewer holds a carved hardwood and polished brass wand, while the room contains thousands of glazed porcelain animals, including peacocks, dogs and cats. The requested scene includes realistic skin and wood textures, ceramic glazing, enamel micro-cracks, dust, film grain, optical lens effects, creaking floorboards, breathing and faint porcelain sounds.

The same sequence includes a real magician beside a tall wooden table with an open rectangular wooden box containing a peacock. It calls for natural lighting, realistic actors and a prestigious modern television-drama aesthetic, without a theater, stage, fantasy or CGI appearance. Another prompt shows a crowned crane walking across snow beneath a clear dusk sky in a documentary style.

Google DeepMind also said it has been experimenting with Gemini 4 Argon and Genie 4 in collaboration with Gemini Robotics. In one example, Argon breaks a user-instructed task into steps for Genie to render, displaying the steps as text captions while the simulation progresses. The companies used Gemini 4 Argon and Genie 4 to simulate Apptronik’s Apollo humanoid playing checkers through real-time task decomposition.

Feedback from trusted testers

Google DeepMind described Genie 4 as an early, experimental model. It has been working with trusted testers in creative industries, journalism, education, gaming and engineering to understand how interactive world models can support real-world workflows and where they need further development.

Creative professionals and technologists are testing Genie 4 as a rapid draft mode for previsualization and ideation. Filmmakers can enter a live simulation and run multiple takes while changing camera angles, lighting, dialogue and action. In this early phase, a live, steerable simulation serves as a shared drafting space in which camera, sound and action can be revised together before any commitment to production.

Photojournalists are exploring how a single two-dimensional archival photograph can become a grounded simulation of its environment. Teams are also testing populated, dynamic worlds for experiencing historical spaces and visualizing short-term events such as flash floods and wildfires.

Game developers and creative technologists are using natural-language prompts to move from an initial idea to an explorable prototype, testing visual styles, materials and interactive mechanics before writing engine code. Enterprise partners in multiple industries are evaluating Genie 4 for rapid zero-to-one scenario generation and identifying gaps in its simulation capabilities.

Toward hybrid world models

Google DeepMind said a promising direction is hybrid world models that combine the structured reasoning of code and language models with the sensory representation of pixel-level simulation. Bringing Genie and Gemini closer together is an early step in that direction.

Gemini can map structure and plan a narrative timeline, while Genie generates the immediate physical reality, including how a world looks, sounds and responds pixel by pixel. Google DeepMind said the combination could produce world models that are more coherent, controllable and grounded than either approach alone.

Safety and transparency

Genie 4 combines real-time generation with photorealism and introduces new capabilities that Google DeepMind said it intends to roll out responsibly. The model includes safety guardrails aligned with the company’s policies for its other models.

Before the preview, Google DeepMind conducted internal safety evaluations intended to prevent harmful or malicious content. It initially limited access to trusted tester partners to gather feedback before a broader release.

All video and audio generated by Genie 4 will include SynthID, Google’s imperceptible digital watermarking technology for identifying AI-generated content across visual and audio formats. Watermarking addresses a core question for photorealistic generation: whether AI-created video and audio can still be identified as synthetic once it circulates.

Current limitations

Genie 4 remains an evolving research project. It can maintain coherent environments during short sessions, but visual details and object permanence may shift over time, particularly during longer simulations, after multiple complex prompt edits or when users navigate extensively through a scene.

Google DeepMind said complex physical interactions, real-time audio accuracy, legible text rendering, spatial accuracy and long-term consistency remain active research areas. These open research areas are what the staged rollout is designed to probe, and they give early users a concrete checklist for judging how far interactive world models have moved from demonstrations toward dependable tools. The company is releasing Genie 4 as a preview to learn with the community while advancing interactive simulation in a controlled manner.

Availability

Genie 4 has been released as a research preview for trusted testers. It is coming to Google AI Ultra subscribers in Project Genie and early-access partners using the Gemini API and Google AI Studio. The waitlist is available through Project Genie.

The original announcement is available on the Google DeepMind Blog.

Acknowledgements

Google DeepMind credited Hang Qi, Diego Rivas, Aleksander Hołyński, Inbar Mosseri, Jack Parker-Holder, Shlomi Fruchter, Tejas Kulkarni, James Assiene, Philip Ball, Jakob Bauer, Eddie Cao, Lucy Chai, Harris Chan, Kleopatra Chatziprimou, Hao Chen, Yoni Choukroun, Mark Collier, Enric Corona, Kurtis David, Ariel Ephrat, Maxime Gazeau, Ruiqi Gao, Chelsee Glover-Odom, Liangke Gui, Swaroop Guntupalli, Shan Han, Tingbo Hou, Jiri Hron, Hui Huang, Drew A. Hudson, Zhouyuan Huo, Linyi Jin, Christos Kaplanis, Dhrupad Karwa, Eunyoung Kim, Congtao Kuang, Eliot Laidlaw, José Lezama, Hao Li, Shuang Li, Yuchen Liu, Hanzi Mao, Mark Matthews, Matt McGill, Kay McKinney, Ricardo Monti, Mark Murphy, Joseph Ortiz, Roni Paiss, Frank Perbet, Ben Poole, Tal Reiss, Konstantinos Rematas, Guy Scully, Jeremy Shar, Shashank Shastry, Sri S, Omer Tov, Ruben Villegas, Andrey Voynov, Hongjie Wang, Luyu Wang, Kaustubh Sridhar, Mo Samsami, Stephen Spencer, Yufei Wang, Will Whitney, Rundi Wu, Joyce (Jingjing) Xie, Jesse Yin, Shiran Zada, Serena Zhang, Xiaofan Zhang, Zhengqi Li, Zhixing Zhang, Yuan Zhong, Shuo-yiin Chang, Dmitry Lagun, Ronny Huang, Yao Su, Ankush Gupta, Genzhi Ye, Haolin Jia, Weiqin Wang, Xi Xiong, Salah Zaiem and Carol Zheng.

The company also thanked cross-functional partners Dina Bashkirova, Emil Bergner, Greta Borisko, Irina Blok, Melissa Byun, Gabe Clapper, Viral Carpenter, Michael Y. Chang, Maximo Cudich-Sieburger, Chloe Doan, Alisa Fortin, Shivani Ghanta, Vikram Gopali, Ezra Gorman, Khyatti Gupta, Sam Hasinoff, Philipp Henzler, Hector Hinestroza, Ed Hirst, Tyler Holland, Kristian Holsheimer, Tiffany Hu, Kapil Joshi, Randeep Katari, Pavan Kumar, Vikhyat Kumeria, Sam Kvaalen, Matthew Levine, Yinxiao Li, Yuchi Liu, Miao Luo, Nick Matarese, Kangfu Mei, Signe Nørly, Moises Oliveira, Eric Paskie, Manika Puri, Susanna Ricco, Andrey Ryabtsev, Cai Schumann, Nicole Brichtova, Andy Sekyere, Grace Shih, Melissa Smith, Duncan Noble Smith, Kenda Stewart, Rod Strougo, Jenn Sturgeon, Amy Zeng, Gaby Pearl, Eleanor Tomlinson, Kenyon Watson, Inês Ayer, Luke Talbot, Han-Yun Tsao, Guillaume Vernade, Karen Wang, Changchang Wu, Ian Wilkinson, Weijuan Xi, Jennifer Yick, Yunhan Zhao, Eric Zhang, Jason Zhang, Annie Zhou, James Zhou, Chris Walker, Alan Yam, Alex Chen, Cliff Lungaretti, Dan Motzenbecker, Kaloyan Kolev, Matthew Carey, Nate Zucker, Sam Lawton, Shashwath Santosh and Suz Chambers.

Google DeepMind also thanked Raia Hadsell, Aäron van den Oord, Josh Woodward, Steve Seitz, Carlos Hernandez, Elias Roman, Jung E. Kim, Abhishek Bapna, Eli Collins, Tulsee Doshi, Carolina Parada, Jie Tan, Koray Kavukcuoglu and Demis Hassabis for their guidance and support, as well as trusted testers and early-access users who provided feedback.