Google DeepMind unveils Genie 4 with native audio

Genie 4 arrives as the fourth generation in a line that began in March 2024 and moved through Genie 2 in December of that year to Genie 3 on August 5, 2025. Genie 3 was introduced as a general-purpose world model that generates diverse interactive environments from text, and it supported real-time operation at 720p and 24fps with visual memory of about one minute. In January 2026, an early-access version called Project Genie was offered to U.S. residents aged 18 and over who subscribe to Google AI Ultra, with sessions limited to 60 seconds of world exploration.

Beyond the audio capability, DeepMind describes Genie 4 as built on the Gemini family: Gemini Omni is said to ground the simulation in real-world appearance and behavior, while Gemini 3.8 Flash serves as a guide agent that observes scenes, divides goals into steps, checks frames, and replans when the world changes or the user intervenes. DeepMind research director Tejas Kulkarni said at a press briefing that "hybrid world models" — in which Gemini writes code and text while Genie simulates reality — are likely to become more common.

This division of labor resembles the plan-act-check loop familiar from AI agents: one side generates video and sound, the other plans and supervises. However, it remains unclear from public information whether Gemini Omni and Gemini 3.8 Flash are embedded inside the Genie model or direct it from outside, or how the components are connected.