> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Google DeepMind Introduces Genie 4 With Native Audio and Gemini-Guided Worlds - URL: https://www.tao.media/google-deepmind-introduces-genie-4-with-native-audio-and-gemini-guided-worlds/ - Published: 2026-10-29T18:09:46.000Z - Updated: 2026-10-29T18:09:47.000Z - Description: The research preview lets users reshape live simulations as Gemini plans events and characters respond in real time. - Author: Bart Hillerich - Tags: Google, AI, News Google DeepMind has introduced [Genie 4](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/?ref=tao.media), an interactive world model that generates explorable environments with native audio and lets users change what happens through text prompts while a simulation is running. The system combines visual generation with Gemini models that help plan events and maintain continuity as scenes evolve. Genie 4 is rolling out as a research preview for trusted testers. DeepMind plans to bring it to Google AI Ultra subscribers through Project Genie and to early access partners through the Gemini API and Google AI Studio. The model extends world simulation beyond generating a moving scene. Users can navigate an environment, introduce an object or character, and hear dialogue and ambient sounds generated alongside the visuals. DeepMind is exploring applications including film pre-visualization, game prototyping, and virtual environments for training and testing robots and AI agents. ![Genie 4 announcement video](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/HVplvENXUAA-sCS.jpg) [Genie 4 announcement video | Google DeepMind](https://x.com/GoogleDeepMind/status/2115114028338962672?ref=tao.media) ## Steering a World While It Runs A world model represents how an environment changes over time. In Genie 4, that takes the form of a simulation generated as the user explores it, with keyboard controls for movement and camera navigation and natural-language prompts for changing the scene. A user can ask for different weather, move the camera, or add a person or object. Gemini adapts those instructions to the surrounding environment and continues monitoring the scene between inputs, helping the simulation progress rather than waiting for each new command. DeepMind's [Genie 4 demonstration](https://x.com/GoogleDeepMind/status/2115114028338962672?ref=tao.media) shows dynamic environments in which people move, react, and speak. The model builds on Gemini Omni for the appearance and behavior of its worlds, while other Gemini models contribute reasoning about how events should unfold. The [capability upgrades](https://x.com/GoogleDeepMind/status/2115114030503166385?ref=tao.media) include more realistic human movement and physical interactions, including collisions and aerodynamics. Demonstrations include a watermelon on a stone plinth and a crowded Japanese railway station set in the 1980s, where passengers move around a train and platform. ![Genie 4 physics demo: watermelon on a stone plinth](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/tG5a6FgXNO0ziuoS.jpg) [Genie 4 physics demo: watermelon on a stone plinth | Google DeepMind](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/?ref=tao.media) Native sound is another addition. Genie 4 is DeepMind's first real-time world model to generate audio directly. Ambient noise, music, and speech are produced with the environment, and characters can respond to user inputs or events through real-time dialogue. A scene can therefore change acoustically as well as visually, rather than relying on a soundtrack added after generation. ![Genie 4 demo: crowded 1980s Japanese train station](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/nE3WzCXnAYCz8jO-.jpg) [Genie 4 demo: crowded 1980s Japanese train station | Google DeepMind](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/?ref=tao.media) ## How Gemini Guides Multi-Step Events Gemini 3.8 Flash acts as a guiding agent inside the simulation loop. As Genie 4 generates the environment, Gemini observes the scene, divides a goal into steps, and reviews incoming frames to check whether each step succeeded. It can then adjust the plan when the world changes or a user intervenes. That feedback loop adds planning to visual generation. Instead of supplying a fixed sequence of instructions in advance, the guiding model can respond to the simulation's current state and choose what should happen next. ![Genie 4 demo: crowned crane walking across snow](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/X4ZFZZZ6CbDQMkdY.jpg) [Genie 4 demo: crowned crane walking across snow | Google DeepMind](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/?ref=tao.media) DeepMind is also experimenting with Gemini 4 Argon and Genie 4 alongside its Gemini Robotics research. In one demonstration, Argon breaks down a task involving Apptronik's Apollo humanoid playing checkers. The steps are supplied as text captions while Genie renders the simulation, allowing the scene to be steered as the task progresses. Argon has already expanded across [Google's consumer and developer products](https://tao.media/google-rolls-out-gemini-4-argon-across-gemini-app-search-antigravity-and-api/?ref=tao.media). In a separate agent experiment, [four Argon agents completed Minecraft's Ender Dragon challenge](https://tao.media/google-deepminds-gemini-4-argon-team-defeats-minecrafts-ender-dragon-in-27-hours/?ref=tao.media) through code and text over more than 27 hours. The Genie work applies step-by-step planning to generated environments, where the model must also track the evolving visual scene. DeepMind sees this combination as an early route toward hybrid world models, in which language and code models handle structure and planning while Genie generates the immediate sensory environment. The research goal is to make simulations more coherent and controllable by combining those capabilities. ## What Testers Are Building Filmmakers and creative technologists are testing Genie 4 for pre-visualization, exploring a scene before committing to production. They can run different takes inside a live simulation and change camera angles, lighting, dialogue, and action as they go. ![Genie 4 demo: concert hall audience](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/am65-PmOmKmXH_h8.jpg) [Genie 4 demo: concert hall audience | Google DeepMind](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/?ref=tao.media) Photojournalists are exploring simulations generated from individual archival photographs, including historical spaces and scenarios involving floods or wildfires. Game developers are testing visual styles and interaction ideas in explorable prototypes before writing engine code, while enterprise partners are evaluating rapid scenario generation for simulation workflows. DeepMind aims to create varied virtual environments for robotics and agent research, in which systems can be trained and tested before acting in the physical world. Other developers are pursuing related approaches, including [Odyssey-3's use of a shared world model across robots, vehicles, drones, and games](https://tao.media/odyssey-unveils-odyssey-3-world-model-for-robots-vehicles-drones-and-games/?ref=tao.media). The consistency of those environments remains a technical constraint. DeepMind reports that visual details and object permanence (whether objects retain their identity and location as a scene changes) can drift during longer sessions, complex prompt edits, or navigation. Complex physical interactions, audio accuracy, legible text, and spatial consistency remain active research areas. ## Preview Access and Content Watermarking Genie 4 underwent internal safety evaluations before the preview and uses guardrails aligned with DeepMind's other models. Generated video and audio will carry SynthID, Google's imperceptible watermark for identifying AI-generated content. Trusted testers are providing feedback ahead of wider access. Google AI Ultra subscribers and developers can join the [Project Genie waitlist](https://forms.gle/WQnSNEM3omF7Uruu7?ref=tao.media), with the planned rollout also covering early access partners in the Gemini API and Google AI Studio.