Google DeepMind has introduced Genie 3 , a new AI system that transforms a short text description or image into an interactive 3D or 2D game environment with 720p graphics at 24 frames per second. Users can explore the generated scene for several minutes, significantly exceeding the limitations of the previous version, Genie 2. The system allows for programming “events on demand,” enabling changes in weather or the addition of new objects during gameplay without restarting the environment.
Using a keyboard, users control a character in the simulated space, and the model maintains stability and detail in visual memory for about a minute. Genie 3 builds on the work of previous models and uses modern video generation methods from the Veo family. The platform already serves as a testing ground for training AI agents capable of performing multi-step tasks in complex virtual spaces.
Access to Genie 3 is open in a research preview format by invitation. Scientists and digital creators are involved in testing to gather feedback and safety suggestions before a wider launch. Early participants note significantly longer playtime and stable geometry but point out simplified physics and a limited action menu compared to classic game engines.
DeepMind positions Genie 3 as part of the “world as a simulator” strategy, complementing other company products, including AlphaZero and Gemini. According to developers, on-demand environment generation without manual asset creation can reduce data and equipment costs and accelerate the development of general artificial intelligence.
