Genie 4
Released Oct 27, 2026
Google DeepMind's real-time interactive world model, which generates explorable worlds from text and image prompts with native audio and lets users steer the scene mid-simulation in natural language, with Gemini models keeping it coherent.
- Output
- Live video stream
- Interactive
- Yes
- Inputs
- text, image, actions
- Max session
- —
- Resolution
- —
- Frame rate
- —
- Access
- Research preview
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Native real-time audio: ambient sound, music and interactive dialogue
- Mid-simulation steering by text, e.g. changing the weather, moving the camera or adding characters
- Text and image prompts, with keyboard navigation and camera control
- Demonstrated as a simulator for Gemini-driven agents and robots (e.g. Apptronik Apollo)
Weaknesses
- Limited research preview for trusted testers; no public access, pricing or API yet
- Google says it supports a few minutes of continuous interaction, with visual drift in longer sessions
- Unreliable text rendering and imperfect simulation of real-world locations
Announced 2026-10-27 (Google blog) as a research preview rolling out to a small group of trusted testers. Successor to Genie 3, which keeps its own entry because Project Genie still runs Genie 3 until the Ultra rollout. TechCrunch reports scenes stay consistent 'up to two minutes, double Genie 3'; Google's pages say only 'a few minutes of continuous interaction', so maxSessionSec is null. No resolution or fps is published for Genie 4 (the 720p/24 fps figures in search results are Genie 3's). Robotics demos pair it with Gemini 3.8 Flash as a guiding agent and Gemini 4 Argon. All output carries SynthID. openWeights false = none mentioned.