DeepMind, Google’s core artificial intelligence (AI) research lab, on Tuesday (October 27) released the latest version of its world model, Genie 4, further enhancing real-time interaction in virtual environments, allowing users to change the weather, adjust camera angles, or add characters and objects through natural language commands while a scene is running.
Unlike traditional AI video generation tools that mainly produce fixed clips based on prompts, Genie 4 can continuously generate subsequent frames as the user interacts, and combines the planning and reasoning capabilities of the Gemini model to explore applications of world simulation technology in game development, film production, and robot training. Here are the three key features of the new model:
1. First-Ever Support for “Native Audio”
The new Genie 4 model can generate ambient noise, background music, and spoken dialogue while building virtual environments, allowing sound to change with events in the scene rather than simply repeating pre-recorded sound effects. For example, a virtual coastline can be accompanied by the sound of waves, and characters can produce corresponding sounds as they move; the model also begins to support interactive dialogue, allowing virtual characters to respond based on user input and scene conditions.
2. Simulation Time Extended to 2 Minutes
Genie 4 can maintain a coherent simulation for about 2 minutes in a single run, double the roughly 1 minute of its predecessor, Genie 3. Users can create a virtual world through text prompts, then use a keyboard to control characters and explore the environment, or issue new commands while the simulation is running. For example, users can ask the system to change the weather, adjust the camera angle, or add new characters and objects to the existing scene; the model modifies the running environment directly, without having to regenerate the video from scratch each time.
3. Gemini Handles AI Training Planning
Another major advance in Genie 4 is the integration of the Gemini model’s planning and reasoning capabilities, enabling AI to not only generate virtual scenes but also arrange actions according to task goals. In this architecture, Gemini 3.8 Flash is responsible for interpreting user goals, breaking down task steps, and observing changes in the generated frames, then guiding Genie 4 to continue the simulation. When the environment or user instructions change, the system can adjust subsequent processes accordingly.
Google is also exploring the use of the larger Gemini 4 Argon model for related research, where the model first plans the task and then hands it over to Genie 4 to simulate the execution process in real time.
This architecture holds promise for robot training, where researchers can first test different manipulation methods in a virtual environment before applying the results to physical robots. Since real-world testing requires equipment, venues, and repeated experiments, realistic simulated environments can help reduce development costs. However, narrowing the gap between simulation and reality remains a major challenge.
Potential Collaboration in Film Production? Genie 4’s Application Potential Comes with Technical Limitations
That said, Genie 4 still faces limitations such as visual drift, inaccurate physics interactions, unstable audio generation, and unclear text recognition, and it remains far from developing a game ready for official release. Google has also admitted that at this stage the model cannot yet replace existing game and film production workflows.
Currently, Google is offering Genie 4 as a research preview to vetted testers, with plans to expand access through Project Genie. The generated video and audio will also carry SynthID digital watermarks to identify AI-generated content. Meanwhile, Google is exploring practical applications with partners in film, gaming, education, and robot simulation; whether it can develop into a mature commercial tool will depend on improvements in the model’s simulation accuracy, stability, and ability to run continuously.
