Google DeepMind launches Genie 3, the first AI that generates interactive 3D worlds in real time

Google DeepMind has unveiled Genie 3, a new frontier for world models, and the first AI that generates interactive 3D worlds in real time at 24 frames per second. The system can generate interactive 3D environments on the fly from nothing, holding together at 720p resolution for minutes at a time. All it needs is a text description. No pre-built 3D assets, no game engine — just words and processing power.

“Today we are announcing Genie 3, a general-purpose world model that can generate an unprecedented diversity of interactive environments. Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p,” Google DeepMind said in a blog post on Tuesday.

Genie’s journey began last December with Genie 2, which introduced 3D world generation and the ability to remember parts of a scene after the user or an AI agent left and returned. That leap addressed one of the biggest weaknesses in prior systems, like Decart’s Oasis, which struggled to recall layouts from earlier in a session.

How Genie 3 works

Genie 3 doesn’t look like a dramatic leap at first glance, but under the hood, it’s a notable upgrade. Genie 3 creates each world one frame at a time, keeping up to a minute of past scenes in memory. Type a prompt, start exploring, and the AI reshapes the environment on the fly as you move.

2/ How does it work? Genie 3 creates worlds frame by frame, remembering up to a minute of past scenes. You enter a text prompt, navigate in real time, and the AI dynamically adapts the world to your movements. pic.twitter.com/RrUSf1WZp1

— Chubby♨️ (@kimmonismus) August 5, 2025

Resolution has doubled from 360p to 720p, and it can keep simulations consistent for several minutes instead of just seconds. Where Genie 2 often started to lose track of its world before hitting the one-minute mark, Genie 3 sustains the illusion much longer before artifacts creep in.

Another major addition is what DeepMind calls “promptable world events.” Genie 2 responded to movement commands, but there was a delay before it generated the next frame. Genie 3 works in real time — and you can now inject new elements into a scene mid-simulation using text commands. In one demo, a skier was joined by a herd of deer at the user’s request. The deer didn’t move perfectly, but the concept opens a new layer of interactivity.

Genie 3: Creating Dynamic Worlds That You Can Navigate in Real-Time

DeepMind still sees Genie 3 as primarily a training ground for AI agents rather than a consumer tool. Fruchter points to scenarios like preparing a self-driving car for “what if” moments that might not appear in its pre-training data. “There are a lot of things that have to happen before a model can be deployed in the real world, but we do see it as a way to more efficiently train models and increase their reliability,” he said.

The tech isn’t flawless. Genie can’t recreate real-world locations with exact fidelity, has trouble rendering text, and can’t yet maintain a persistent simulation for hours. But Parker-Holder says it’s already proving useful: “You can certainly find things you wouldn’t want agents to do because if they act unsafe in some settings, even if those settings aren’t perfect, it’s still good to know. You can already see where this is going.”

For now, Genie 3 isn’t available to the public, though DeepMind plans to expand testing.

Under the hood, Genie 3 works by generating each frame based on the full history of the session. At 24 frames per second, that’s 24 separate calculations every second, each referencing minutes of prior data. In practice, that makes it possible to explore, alter, and adapt a simulated world in real time.

DeepMind has already run its SIMA agent through Genie 3’s worlds, pushing longer action sequences and more complex goals than before. The implications stretch beyond gaming — robotics training, autonomous navigation, and AI safety research could all benefit from richer, more responsive simulated environments.

Below is a YouTube video of Genie 3 in action.