Genie 4 launches as DeepMind's first world model with native audio
Genie 4 adds real-time ambient audio and character dialogue, holds scenes for about two minutes, runs on Gemini 3.8 Flash, and pairs with Gemini 4 Argon for robot training.
Genie 4 is out and it's the first DeepMind world model with actual audio baked in. Not slapped on after generated alongside the visuals. Ambient sound, music, characters talking back to you in real time. Scenes now hold for ~2 minutes, double Genie 3. Gemini 3.8 Flash runs in the loop as the "brain," keeping causality and continuity straight while you reshape the world mid-scene. Change the weather, swap camera angles, drop in a new character, and it adapts without restarting. Demos so far: a saxophonist on a floating lake platform in the fog, silk drifting through a wildflower meadow, a hot air balloon over Antarctica, a packed 1980s Japanese train station. All photoreal, all reactive. This isn't a toy. DeepMind is framing it as training ground for robots and agents simulate the task before you risk the real world. Pairs with Gemini 4 Argon for stuff like a humanoid robot working through checkers moves or pipette tasks step by step. Currently a research preview, trusted testers only. Coming to Google AI Ultra subscribers through Project Genie and to partners via the Gemini API. Waitlist's open. Still drifts on long sessions, still struggles with legible text and multi-agent interactions. But going from silent 2D-ish worlds to synced audio-visual sims with memory in three generations is not a small jump. The world model race just got a soundtrack.