Genie 4

Canonical version: Genie 4.

Genie 4 is a world model from Google DeepMind, announced on 27 October 2026 by Raia Hadsell, Diego Rivas and Hang Qi. You describe a world in text, and Genie generates it as video you can move through and change in real time. Think of it as a game engine where the "rendering" is a neural network that predicts the next frame from what you do.

A world model is a model that learns how an environment behaves (what happens next when you act) instead of just producing text or a fixed clip. That's the difference with video generators like Gemini Omni, Veo 3 or Sora: those render a clip and stop, while Genie keeps the world running and reacts to your input.

What's new

• Native audio. DeepMind calls it its first real-time world model that generates audio natively: music, ambient sound and dialogue
• Change the world by talking to it. Ask for rain, move the camera, add a person or an object, and the scene adapts. Gemini rewrites your request to fit the scene
• Physics and characters. Aerodynamics, collisions, and characters that react to their surroundings and hold conversations that adapt to what you do
• A Gemini agent inside. Gemini 3.8 Flash can act as a guiding agent: it watches the scene, splits a goal into steps and checks each step
• Built on Gemini. Gemini Omni grounds how things look and move; other Gemini models handle the reasoning that keeps the scene coherent
• SynthID watermark on all generated video and audio

DeepMind published no numbers this time: no resolution, frame rate, latency or session length. For comparison, Genie 3 (August 2025) ran at 720p and 24 fps, stayed consistent for a few minutes and remembered about one minute back.

Who gets it

It's a research preview for trusted testers. It rolls out first to Google AI Ultra subscribers in the Project Genie app, and to early access partners of the Gemini API and Google AI Studio. Everyone else joins a waitlist. No pricing, no country list.

What people use it for

DeepMind's examples:

• Film: run several "takes" of a scene live, changing camera, light, dialogue and action
• Games: go from an idea to an explorable prototype, try visual styles and mechanics
• History and training: turn one archival 2D photo into a simulation; rehearse flash floods or wildfires
• Robotics: a demo shows Apptronik's Apollo humanoid playing checkers

Notice what moved. Genie started as a way to train agents in endless generated worlds (Genie 3's pitch was "what if" scenarios for agents). Genie 4's demos are mostly about film, games and enterprise use. ainch made that point on Hacker News; roryirvine read it as a sharper focus rather than a change of plan.

Limits

DeepMind lists them itself: visual details and object permanence still drift over time, and legible text, spatial accuracy, long-term consistency, complex physical interactions and real-time audio accuracy are all open research.

The HN thread (1,044 points) found them fast in the demo videos:

• A second castle appears after the camera turns away and comes back; a rider's helmet changes; a wall in the Van Gogh bridge example doesn't match itself
• A prompt asking for photorealism came out looking like a video game, and a saxophone prompt produced a clarinet
• Several commenters suspect the demos are cherry-picked

The core argument (cvoss, edstarch): Genie only "knows" what's on screen. Anything outside the frame has to be re-imagined when you look back, so permanence will stay fragile without some persistent world state. Proposed fixes ranged from tracking objects and camera position, to rendering a low-fidelity 3D scene first, to training on autonomous-vehicle data (360° cameras plus lidar) instead of YouTube.

DeepMind seems to agree. Its own "next step" is hybrid world models: Gemini maps the structure and plans the story, and Genie renders the world and its reactions.

My take

The hybrid direction is the interesting part. A pure next-frame predictor can't keep a castle in place once you look away, and a separate planner that holds state is the obvious fix. That's the same pattern we use for agents: the model generates, something else keeps the memory.

For me it's still a demo, not a tool. Ultra-only, waitlisted, no numbers. But the business logic I noted in FLUX.3 Mimic applies here too: media generation pays for the world model, and the world model is what robotics needs. The checkers-playing humanoid is the tell.

Genie 3

Announced 5 August 2025 as a limited research preview for a small group of academics and creators. 720p at 24 fps in real time, consistent for a few minutes, visual memory of about one minute, and "promptable world events" (change the weather, add objects or characters). Its stated limits: a small action space for agents, weak multi-agent interactions, inexact real-world geography, and sessions of minutes rather than hours.

References

• Google DeepMind, "Genie 4: A new generation of world simulation" (2026-10-27): https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/
• Hacker News discussion: https://news.ycombinator.com/item?id=50261319
• Google DeepMind, "Genie 3: A new frontier for world models" (2025-08-05): https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

Related

• Google DeepMind
• Gemini
• Gemini Omni
• Veo 3
• Sora
• FLUX.3 Mimic
• MuZero
• AI Agents
• Google AI Studio
• Generative AI (Gen AI)

About Sébastien

Ready to get to the next level?