Most AI video tools generate a clip, and if you want sound, a second model bolts it on afterward — a soundtrack guessed from frames that already exist. Genie 4, which Google DeepMind announced on October 27, generates the sound with the world, not after it: the company calls it “our first real-time world model to generate audio natively,” meaning ambient sound, music and character dialogue all come out of the same simulation that’s rendering what you see.
That’s the headline difference from Genie 3, the predecessor. Genie 4 is an interactive world model: point it at a prompt, a sketch, or a single archival photo, and it builds an explorable, persistent environment you move through with a keyboard, steering events with plain-language instructions — change the weather, add a person, move the camera. DeepMind says physics and human movement are more convincingly simulated this time, pointing to scenarios like aerodynamics and collisions, and characters that react to what’s happening around them rather than idling in place.
Gemini does the thinking Genie doesn’t. Gemini Omni grounds each scene in how real places and objects actually look and behave, while other Gemini models handle the reasoning needed to keep a simulation coherent over time. For more complex, multi-step events, Gemini 3.8 Flash acts as a guiding agent — breaking a goal into steps, checking progress, adjusting the plan as the world changes underneath it. DeepMind also connects the work to its robotics line: an example pairs Gemini 4 Argon with Genie 4 and Gemini Robotics to have Apptronik’s Apollo humanoid play checkers, using the simulation to plan physical action rather than just depict it.
What it’s actually for, according to testers: previsualization, turning a single archival photo into an explorable simulation, prototyping game concepts before building them, and generating scenario variations for enterprise use. None of that is rendered final-pixel content — it’s exploratory and disposable, the same way a sketch is disposable in a way a painting isn’t.
The limits are the same ones every world model has had so far, just pushed back a little. DeepMind is explicit that visual details and object permanence can drift over longer sessions or after many edits — a chair you walked past might not be the same chair when you turn around. Complex physics, accurate audio, legible text and long-term spatial consistency are all still described as active research, not solved problems. Every generated frame and audio clip carries a SynthID watermark, and the company says the model went through internal safety evaluation before this preview.
Access is narrow for now: a research preview for trusted testers, with a waitlist open for Google AI Ultra subscribers through the Project Genie web app and for early-access partners on the Gemini API and Google AI Studio. DeepMind frames the long-term goal as “hybrid world models” — combining Gemini’s structured reasoning with Genie’s pixel-level rendering into one system, rather than treating narrative logic and visual simulation as separate problems.
Related Reading
- Genie 4: A new generation of world simulation — Google DeepMind
- Google DeepMind’s Genie 4 puts a modern spin on agent designs that predate LLMs — The Decoder
- Meet Google DeepMind’s Genie 4: Real-Time World Model With Native Audio — MarkTechPost
- DeepMind unveils Genie 4 world model promising more realistic physics and humans — Ars Technica
- SynthID — Google’s content provenance watermarking