Real-Time Interactive Simulation
On 27 October 2026, Google DeepMind announced Genie 4. The company described the new model as a world model. It creates interactive simulations in real time. The simulation is not a pre-recorded video. As the user acts, the world responds. Image, motion and sound take shape in a single process. DeepMind said the new model's level of realism is higher than in the previous stage. Here, realism means not only that the appearance looks close to a photograph, but also that the world moves coherently. A scene does not freeze in a single frame. As events unfold, the world develops with them. The user controls the scene not as an observer, but as a participant in the process. This is what sets Genie 4 apart from ordinary video generation. The model's task is not to create a single image or a single clip. Its task is to create a world that continues, changes and responds. The term interactive simulation carries exactly this meaning: the world changes in response to the user's influence, and the change appears on screen immediately.
Control in Natural Language
Genie 4 scenes can be controlled in real time in natural language. The user can move the camera. The user can change the weather. The user can introduce a new character into the scene. Every command is given as text. The Gemini model adapts the command to the surrounding scene. For example, if the scene is taking place in the evening by the sea, a new character appears in a way that fits that setting. A command is not detached from the scene. It is carried out within the logic of the scene. This approach preserves the continuity of the simulation. After the user gives a single command, the world does not stop. The world moves to a new state on the basis of the command and continues in that state. Camera movement is part of the same process. A change in the weather means not only a change in the scene's colour, but also a change in its overall mood and lighting. When a new character enters the scene, the world's other elements respond to it as well. In this way, text commands become control over a living world.
Native Audio and Dialogue
Genie 4 is a native multimodal model. The model creates music. It creates ambient sounds. It creates interactive dialogue. All three kinds of sound are generated at the same time as the simulated world. Sound is not added separately after the image is ready. It emerges as part of the world. The sound of wind, water or city noise matches the setting of the scene. When characters speak, dialogue is also formed in real time. Dialogue is not pre-written text. It is created in response to events in the scene and to the user's influence. DeepMind described Genie 4 as the first real-time world model to generate audio natively. This description marks a change in the model's history: sound is no longer an added layer, but an internal part of the model. At the same time, all video and audio created by Genie 4 is marked with a SynthID watermark. The SynthID mark helps identify content created with the help of artificial intelligence. The marking applies to both video and audio.
Built on Gemini Models
Genie 4 runs on the Gemini family of models. Gemini Omni grounds simulations in real-world appearance and behaviour. Other Gemini models perform the reasoning needed to preserve the coherence and duration of a simulation. This division shows two sides of the model. On one side are how the world looks and how characters move. On the other side are the logic and sequence of events. Without the first side, the world would look unconvincing. Without the second, the world would quickly fall apart and the logic of the scene would break. By grounding itself in real-world appearance, Gemini Omni gives the simulation a visual foundation. Through reasoning, other Gemini models follow the development of the scene and preserve the connections between events. As a result, a scene lives not for a few seconds, but as a continuing process. A character's movement matches its surroundings. A change in the surroundings affects the character's movement. This mutual interaction is the central mechanism of Genie 4's simulation.
"Today we are announcing Genie 4. This is our new world model. It unlocks an unprecedented level of realism and advances real-time interactive simulation." — Google DeepMind blog
Testing Phase and Rollout Order
Genie 4 is currently being released in a research preview phase. In the first phase, the model will be open to trusted testers. Feedback will be gathered before a wider release. In the next phase, the model is said to open to Google AI Ultra subscribers as part of Project Genie. Early-access partners through the Gemini API and Google AI Studio will also use the model in the next phase. The rollout order is set out in stages. First, a limited test group will work with it. Then subscribers and partners will join. On 27 October 2026, The Verge also published an article about Genie 4. The publication confirmed the announcement as the day's main DeepMind news. Two independent sources — the official DeepMind blog and the Google blog — described the announcement, the model's main features and the rollout order in the same way. The Verge article additionally confirmed that the announcement came out on that day. The Google DeepMind blog and the Google blog are the main sources for the announcement.
