Artificial Intelligence
Google Deepmind unveils Genie 4 world model
Google Deepmind introduced Genie 4, a real-time world model that simulates interactive 3D scenes with sound and dialogue, while Gemini models handle planning and the system remains experimental.
- Google Deepmind introduced Genie 4 as the successor to Genie 3, which it had shown in August 2025.
- The model simulates scenes frame by frame and lets users move through them with keyboard controls and text prompts.
- Deepmind said Genie 4 generates background audio and character dialogue directly.
- The company said longer sessions can produce visual drift and missing objects, and access will begin with select testers before expanding to Google AI Ultra subscribers.
Google Deepmind said Genie 4 was its successor to Genie 3, which it had presented in August 2025. It described the system as a world model that updates a scene one frame at a time instead of outputting a finished video.
Users can move through the simulation with keyboard controls and influence it with text prompts. Deepmind said the model also creates background sound and spoken exchanges for characters inside the scene.
The company said Gemini models operate behind the scenes to keep events coherent. It said Gemini 3.8 Flash watches the scene, breaks goals into steps and checks whether each step worked, while Gemini 4 Argon was used in a demo that simulated Apptronik's Apollo robot playing checkers.
Deepmind said Genie 4 remained an experimental preview with limits in longer runs, including changes in visual detail, missing objects and unresolved physical interactions. It said access would begin with select testers, then expand to Google AI Ultra subscribers in Project Genie, with early access also planned through the Gemini API and Google AI Studio, and that SynthID would mark generated video and audio.
Named in this story
Companies
- Google Deepmind
- introduced Genie 4
- Apptronik
- made the Apollo humanoid robot
- said the generated videos and audio would carry its watermark
Products and systems
- Genie 4
- the world model being announced
- Genie 3
- the predecessor model shown in August 2025
- Gemini models
- planned events in the background
- Gemini Omni
- provided video capabilities that Genie 4 built on
- Gemini 3.8 Flash
- watched scenes and checked multistep progress
- Gemini 4 Argon
- was used in a robotics demo
- Gemini Robotics
- the robotics line Deepmind linked to Genie work
- Apollo
- was simulated playing checkers
- Project Genie
- the web app where access was to expand
- Google AI Ultra
- the subscriber tier that would get access later
- Google AI Studio
- another channel for early access
- SynthID
- the watermarking system Google said would mark generated media
How the source tells it
A mostly descriptive product announcement, with light novelty and vendor-claim framing.
- novelty the piece emphasises first-time capabilities and a successor model to signal a new step
- vendor boosterism Deepmind's own claims about realism, planning and future uses are repeated with little independent checking