What is Genie 4?
A world model, not a video generator. Announced October 27, 2026, Genie 4 builds interactive environments you steer with language — with physics, audio, and characters that respond.
Definition in one paragraph
Genie 4 is Google DeepMind's real-time, natively multimodal world model, announced on October 27, 2026. Given a description, an image, or a video, it generates an explorable simulated environment that keeps running and responds to your instructions — adding weather, moving the camera, introducing characters, or triggering events — without restarting. Audio is generated with the scene rather than added later.
- It is generative — Worlds are synthesised from a prompt or a starting image, not authored in an engine.
- It is interactive — The scene reacts to your input in real time instead of playing back a fixed render.
- It is multimodal — Visuals, ambience, music-like sound, and dialogue are produced together.
- It is agent-driven — Gemini models reason about goals and check their own output inside a live loop.
The name, unpacked
DeepMind's own framing calls Genie 4 "a new generation of world simulation." The word choice matters: simulation implies state, rules, and causality, not just pixels arranged plausibly frame by frame.
World
A persistent environment with geometry, lighting, and consistent identity over time.
Model
A learned system that predicts the next state of that world, not a stored asset.
Interaction
Your input is part of the prediction loop, so the world branches as you act.
Real time
Output is recomputed continuously; you do not wait for a batch render.
How the Gemini-based system is put together
DeepMind describes Genie 4 as a composition of Gemini models, each doing a different job. This is why it can hold a scene together across edits — and why its errors look like reasoning errors rather than rendering artefacts.
Genie 4 versus a text-to-video model
The distinction determines what you can actually do with it, so it is worth being explicit.
| Question | Genie 4 (world model) | Text-to-video model |
|---|---|---|
| What you get | A running, explorable environment | A rendered clip with fixed frames |
| Can you change it mid-way? | Yes — weather, camera, objects, events | No — you regenerate from the start |
| Duration | An open session you end when you choose | A fixed length decided up front |
| Physics | Modelled as world state that persists | Inferred per frame, may break on contact |
| Audio | Native ambience and reactive dialogue | Usually added in post |
| Best use | Agent training, prototyping, pre-visualisation | Finished short clips and b-roll |
Scroll the table sideways to see all columns.
What Genie 4 does not do well
DeepMind states these limitations directly. Repeating them accurately is part of using the model responsibly.
Drift
Visual detail and object permanence degrade over long sessions or after many edits.
Complex physics
Unusual interactions are where the simulation is most likely to break its own rules.
Audio accuracy
Sound is generated natively but is not guaranteed to be acoustically correct.
Legible text
On-screen writing is unreliable unless it is explicitly part of the prompt.
Spatial precision
It cannot reproduce a real location with perfect geographic accuracy.
Long-term consistency
Sessions are not designed for hours of continuous exploration.
Quick answers
The definition-level questions, answered.
Now see what it can do
The capabilities page walks through real-time steering, physics, native audio, and complex event simulation — with the exact claims Google makes.