Google DeepMind has unveiled Genie 4, an AI model designed to generate interactive virtual worlds in real time. Unlike standard video generators, this system models environmental changes based on user input, including keyboard navigation and text commands to modify scenes.
The model now supports simultaneous audio generation, including ambient noise, music, and dialogue. According to TechCrunch, Genie 4 maintains scene consistency for up to two minutes, doubling the duration of its predecessor. It also features improved physics simulation, particularly for human movement and aerodynamic interactions.
Genie 4 utilizes the Gemini model family, with Gemini 3.8 Flash acting as an agent to manage complex, multi-step tasks. Google is also testing the model alongside Gemini 4 Argon within the Gemini Robotics research program to simulate tasks for robots, such as the Apptronik Apollo.
The model is currently available to a limited group of testers as a research preview. Access will expand to Google AI Ultra subscribers, Project Genie users, and partners via Gemini API and Google AI Studio in the coming weeks. A waitlist for Project Genie is currently open.
Google is applying SynthID digital watermarks to all audio and video content generated by the model. While the company is targeting applications in film, gaming, and education, it notes that Genie 4 remains an experimental model with ongoing work needed on spatial accuracy and long-term object consistency.
