Google DeepMind has released Genie 4 as a research preview for trusted testers — its first real-time world model with native audio, live dialogue and text steering built in. The model generates explorable worlds from text and image prompts in real time, complete with ambient sound, music and characters that talk back.
Google DeepMind Genie 4 (Credits: MarkTechPost)
Genie 4 is not a video clip generator. It simulates interactive environments frame by frame, with each frame responding to user input as it arrives. Users move with keyboard controls and steer events with plain-language prompts — changing the weather, moving the camera, or adding a new person or object mid-session — while the world keeps progressing between inputs.
Google DeepMind Genie 4: Gemini Inside the Simulation Loop
The bigger change is architectural. A Gemini model now runs inside the simulation loop as a planner. Gemini 3.8 Flash breaks a goal into a sequence of steps, Gemini Omni anchors the look and behaviour of each world, and other Gemini models handle the reasoning that keeps a simulation coherent — interpreting prompts to fit the running scene, verifying each step against incoming frames, and revising the plan when the world shifts or the user intervenes.
DeepMind frames this as an early step toward hybrid world models, in which a language model handles structure and narrative while Genie renders pixels and sound. Scene consistency now stretches to about two minutes, double Genie 3’s one minute of visual memory, according to TechCrunch’s briefing — though sessions still last only a few minutes, and objects can drift after repeated prompt edits.
Genie 4 vs Genie 3: What’s New
Genie 4 succeeds Genie 3, announced in August 2025, which ran at 720p and 24 frames per second. Genie 4’s resolution and frame rate have not been disclosed, but its headline advances are clear: it is DeepMind’s first real-time world model to generate audio natively, with every world carrying ambient audio and characters able to hold conversations that react to inputs in real time. DeepMind also cites physics simulation — aerodynamics and collisions — and calls human motion its biggest leap, with characters that react to their surroundings.
One research demo shows the practical ambition: Gemini 4 Argon split a checkers task into steps, and Genie 4 received each step as a text caption and simulated Apptronik’s Apollo humanoid playing the game. That points to the target audience — anyone building training environments for agents or robots, where a world that responds, talks and keeps its own story straight is far more useful than a pretty video.
Access remains limited for now: Genie 4 is Google-hosted only, live for trusted testers via the Project Genie web app, with availability expected soon for Google AI Ultra subscribers in Project Genie, followed by early access partners of the Gemini API and Google AI Studio. DeepMind has released no benchmark tables for the model, so independent testing will be the real verdict on how far this hybrid approach has come.
Also Read: Google’s Genie 4 World Model Adds Real-Time Audio, Photorealistic Worlds and Mid-Simulation Steering FlashMap: Polymathic AI and Princeton’s Flow Map Drafter Delivers 6.14x Faster LLM Decoding
Source: MarkTechPost — read the full story there.
