Skip to content

Back to the catalog

deepmind.googlearticle

Genie 4: A new generation of world simulation

Genie 4 is Google DeepMind's real-time interactive world model, announced as a research preview for trusted testers. It generates video with native audio and interactive dialogue, steered by keyboard controls and text prompts. Gemini 3.8 Flash acts as a guiding agent that plans steps and checks frames, and Gemini 4 Argon is used for robotics experiments. The post lists limitations such as drift and object permanence, and says outputs carry SynthID watermarks.

TL;DR Announces Genie 4, a real-time world model with native audio and dialogue, guided by Gemini models, in research preview.
Open original ↗
Document

What it claims

Genie 4 is a Google DeepMind world model that generates interactive video with native audio, including ambient sound, music and character dialogue, in real time. Users steer it with keyboard controls for navigation and camera, and with text prompts that change weather, camera or add people and objects mid-session. DeepMind says the biggest gain is in simulating people who move, react and speak. The post gives no benchmarks, model size, latency, resolution or session-length figures.

How it works

Genie 4 draws on several Gemini models. Gemini Omni grounds the look and behaviour of the simulation in the real world. Other Gemini models handle the reasoning that keeps the scene coherent. Gemini rewrites each user prompt to fit the current scene and steers the world between user inputs.

For multi-step events, Gemini 3.8 Flash runs as a guiding agent in a closed loop. It splits a goal into steps, sends each step to Genie 4, reviews incoming frames to check that the step succeeded, and replans when the world changes or the user intervenes.

flowchart LR
  G[User goal] --> P[Gemini 3.8 Flash plans steps]
  P --> R[Genie 4 renders frames]
  R --> V{Step verified?}
  V -- yes --> P
  V -- no / user intervenes --> P

In robotics experiments, Gemini 4 Argon splits a task into steps and feeds them to Genie 4 as interleaved text captions while the simulation renders. DeepMind describes this as a path toward "hybrid world models": language and code models plan the structure and timeline, and Genie renders pixels and sound.

How to use it

Genie 4 is a research preview for trusted testers only. Access is planned first for Google AI Ultra subscribers through Project Genie and for early access partners of the Gemini API and Google AI Studio. Project Genie has a waitlist. The post documents no API shape, pricing or weights. Testers have used it for film pre-visualization, turning a single 2D archival photo into a simulated scene, game concept prototyping, and generating scenarios for enterprise simulation.

Limits

DeepMind calls Genie 4 early and experimental. Scenes stay coherent only over short sessions. Visual details and object permanence drift as sessions run longer, after several complex prompt edits, or as the user navigates through the scene. The post lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as unsolved. All generated video and audio carries a SynthID watermark, and the model has safety guardrails that follow DeepMind's policies for its other models.

Organisations named here

Technologies named here

Techniques named here

  • Gemini as guiding agent in the loop
  • Real-time text steering of world simulation
  • Interleaved text captions for task execution
  • Hybrid world models

Related in models