Google DeepMind previews Genie 4 interactive world model

Google DeepMind has released Genie 4 as a research preview, introducing a world model that generates interactive, more photorealistic simulated environments with synchronized real-time audio. The company says the model can keep a scene consistent for up to two minutes, twice the limit of its Genie 3 predecessor released a little more than a year ago.

Genie 4 allows users to enter text prompts while a simulation is running and change elements of the scene in real time. Google cited edits such as changing the weather or camera angle, or introducing new people and objects. The system also produces music, ambient sounds and interactive dialogue intended to match the generated world.

More direct control over simulated environments

Google positions Genie 4 as an advance in both visual fidelity and interaction. Alongside freeform steering, it offers improved physics and more human-like characters. These capabilities are central to world models, which seek to create environments that react dynamically to a user's actions rather than produce a fixed video sequence.

The work builds on applications such as DeepMind street simulations for robots, where simulated street environments can support robotics scenarios, while Genie 4 expands the interaction model with live edits and audio. Google has also named gaming, interactive learning and robotics among the potential commercial uses for this class of system.

Competition in the category is intensifying. Runway recently released GWM Worlds 2, a playable world model with dynamic responses to user input and audio generation. The field is also seeing consolidation: AMD is acquiring World Labs, Amazon is reportedly in talks to acquire Decart, and Rivian acquired Third Dimension for self-driving simulation capabilities.

Gemini acts as a planning layer

Google is more closely connecting Genie 4 with Gemini. In a research experiment linked to Gemini Robotics, a Gemini 4 Argon model decomposed tasks into steps that Genie 4 simulated in real time. Tejas Kulkarni, research director at Google, described a future of hybrid world models in which Gemini writes code and text while Genie ingests that material and simulates reality at different levels of abstraction.

This arrangement could provide a way to test robotics tasks in simulation before they are attempted in the physical world. Google has also highlighted creative workflows: DeepMind engineer Diego Rivas said rapid, real-time interaction may be useful during ideation and storyboarding. The company already offers Veo for high-definition video generation with synchronized audio, while Rivas said turning interactive prototypes into creative assets with Gemini Omni remains some way off.

Known limitations and availability

DeepMind stressed that Genie 4 remains early-stage technology and that the final end-to-end product experience has not been determined. The team acknowledged that object details can drift as a simulation runs longer. It is also still working on audio accuracy and making in-world text reliably legible.

Genie 4 is initially available to trusted testers. In coming weeks, Google plans to extend access through Project Genie to Google AI Ultra subscribers and to early-access partners using the Gemini API and Google AI Studio. Businesses assessing world-model tools should begin with tightly bounded prototypes, measure consistency and output quality for their own scenarios, and retain human review before relying on simulations for creative or robotics decisions.