Google DeepMind's breakthrough foundation world model – The first real-time interactive general-purpose world model. Creating dynamic 3D environments at 720p resolution, 24fps with world memory up to 1 minute. Built on Veo 3's physics understanding and self-supervised learning. A crucial stepping stone toward AGI.
Genie 3 is the first real-time interactive general-purpose world model developed by Google DeepMind. Released on August 5, 2025, it transforms text prompts into fully interactive, consistent 3D environments that you can navigate in real time at 720p resolution and 24fps.
Building on Genie 2 and DeepMind's video generation model Veo 3, Genie 3 uses self-supervised learning to teach itself physics, rather than relying on hard-coded physics engines. A key innovation is world memory – environments retain changes you make, with visual memory extending up to one minute for physically coherent long-horizon generation.
"Genie 3 is the first real-time interactive general-purpose world model. World models are key on the path to AGI, specifically for embodied agents, where simulating real world scenarios is particularly challenging."
- Shlomi Fruchter, Research Director at Google DeepMind
Real-time recordings from Google DeepMind showing Genie 3 generating interactive worlds from text prompts
Click anywhere to enable autoplay
Prompt: First person perspective navigating difficult terr...
Click anywhere to enable autoplay
Prompt: Jetski during the festival of lights
Click anywhere to enable autoplay
Prompt: Running by the shores of a glacial lake, exploring...
Click anywhere to enable autoplay
Prompt: An adorable, fluffy creature bounding across a vib...
Videos courtesy of Google DeepMind
Understanding the technical innovations behind Genie 3
Generates one frame at a time based on user input, managing growing trajectories while processing new inputs multiple times per second for real-time responsiveness.
Environments retain changes you introduce. Return to an area and find it exactly as you left it. Visual memory extends up to one minute for consistent generation.
Unlike hard-coded physics engines, Genie 3 learns how physics works through self-supervised learning, discovering patterns from unlabeled data.
Pushing the boundaries of AI world modeling technology
See how Genie 3 advances beyond its predecessor
| Feature | Genie 2 | Genie 3 |
|---|---|---|
| Resolution | 360p | 720p |
| Duration | 10-20 seconds | Multiple minutes |
| Frame Rate | Variable | 24 FPS |
| Promptable Events | Limited | Full Support |
| Memory Consistency | Basic | Up to 1 minute |
Experience the cutting-edge capabilities that make Genie 3 a game-changer in AI world modeling
Generate interactive worlds at 24 frames per second with instant response to text prompts
Maintain environmental coherence for several minutes with auto-regressive frame generation
High-quality visual output at 720p resolution for detailed and immersive environments
Accurately model physical properties and natural laws within generated environments
Dynamically modify environments with text prompts to create custom world events
Developed with safety and ethics in mind, featuring controlled access and responsible deployment
Everything you need to know about Genie 3
Genie 3 was announced on August 5, 2025 and is currently available as a limited research preview to select academics and creators. Google DeepMind plans to expand access gradually, but no public release date has been announced yet.
World memory is a key innovation in Genie 3. Generated environments retain changes introduced by the user – if you alter an object or leave a mark, returning to that area shows the environment unchanged since your last interaction. Visual memory extends up to one minute.
Current limitations include: restricted action space for direct agent control, challenges modeling complex multi-agent interactions, incomplete geographic accuracy for real-world locations, text rendering quality issues, and interaction duration capped at several minutes. Physics inaccuracies and occasional visual hallucinations may also occur.
DeepMind believes world models like Genie 3 are crucial stepping stones toward AGI. They enable training AI agents in unlimited simulation environments, forcing them to learn from experience similar to how humans learn. Genie 3 has been tested with SIMA agents executing complex multi-step goals.
Unlike Veo 3 and other AI video generators that create pre-determined videos, Genie 3 is a live simulator allowing real-time interaction. It generates frames based on user input using auto-regressive synthesis, and crucially, it uses self-supervised learning to understand physics rather than hard-coded rules.
Genie 3 can generate a wide variety of interactive environments including: physical phenomena (water, lighting, environmental interactions), ecosystems with animal behaviors, fantastical and animated scenarios, historical location recreations, and complex terrain navigation scenarios.
Promptable events allow you to dynamically modify generated environments through text prompts. You can change weather conditions, introduce new objects, place characters, and create "what-if" counterfactual scenarios – all in real-time while exploring the generated world.
Unlock new possibilities across multiple domains with Genie 3's world modeling capabilities
Create diverse training environments for AI agents and autonomous systems to learn and adapt in controlled scenarios
Train robots and autonomous systems in simulated environments before real-world deployment
Generate interactive educational scenarios for immersive learning experiences across various subjects
Amplify human creativity by generating animated and fictional scenarios for entertainment and media
Limited spots available for academic researchers and creative professionals. Join the waitlist to get early access to Genie 3's world-building capabilities.