Meet Google DeepMind’s Genie 4: Real-Time World Model With Native Audio
MarkTechPost Michal Sutter ● Covered by 9 sources
Google DeepMind just showed Genie 4, a world model that makes scenes and sound in real time. It’s more than a video toy: Gemini is now inside the loop, steering what happens next.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has put Genie 4 out as a research preview for trusted testers, and the pitch is bigger than “AI makes a fake world.” The model builds explorable scenes from text and image prompts in real time, while also generating ambient sound, music and character dialogue. Users can move around with keyboard controls and change the scene with plain-language prompts.
The real shift is under the hood. A Gemini model now sits inside the simulation loop as a planner, which DeepMind says matters for training environments for agents and robots. In the setup it describes, Gemini first adapts the user’s prompt to the active scene, then breaks the goal into steps, checks incoming frames to see whether those steps worked, and revises the plan when the world changes or the user intervenes.
DeepMind says Genie 4 is a natively multimodal world model, not a video generator pretending to be interactive. It simulates frame by frame, so each frame responds to input as it comes in. The company also says this is its first real-time world model to generate audio natively, and that Genie 3 came earlier, in August 2025.
The demo cases are broad. Users can change weather, move the camera, or add people and objects mid-session. DeepMind also points to aerodynamics, collisions and human motion, with characters reacting to their surroundings. In one research demo, Gemini 4 Argon split a checkers task into steps, and Genie 4 simulated Apptronik’s Apollo humanoid carrying out the game from those text captions.
But the preview comes with tight limits. DeepMind says sessions last only a few minutes, and visual details can drift after repeated prompt edits or longer runs. Real-world locations are not perfectly reproduced, legible text usually appears only when it is already in the prompt, and real-time audio accuracy plus multi-agent behavior are still unsolved. Access starts with trusted testers, then Google AI Ultra subscribers in Project Genie, followed by Gemini API and Google AI Studio partners.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of flex: less fireworks, more plumbing. The interesting part isn’t the shiny world demo, it’s Gemini being used as the planner inside the loop, because that’s where these systems start looking useful for actual simulation work. The rest is familiar AI theater with a nicer soundtrack, which is fine as long as nobody confuses a few-minute preview for a finished world.
Read more about this at: MarkTechPost