Google DeepMind's Genie 4 Generates Real-Time Worlds With Audio

Genie 4 also generates a soundscape, including ambient noise, music, and character dialogue, making it DeepMind's first real-time world model to natively produce audio.

Users interact with the generated worlds using keyboard controls and steer events with plain-language prompts, with the model's architectural change featuring a Gemini model running inside the simulation loop as a planner.

This architectural pattern is noted as important for those building training environments for agents or robots.

Genie 4 offers photorealistic worlds with native audio and live text steering, maintaining scene consistency for up to 2 minutes, which is double that of Genie 3.

However, sessions typically last only a few minutes, and objects may drift after repeated prompt edits.

The model is natively multimodal, generating audio, music, and dialogue concurrently with visuals, succeeding Genie 3, which was announced in August 2025 and ran at 720p and 24 frames per second with visual memory reaching back 1 minute.

Genie 4 integrates two parts of the Gemini family: Gemini Omni anchors the world's look and behavior, while other Gemini models manage the reasoning to keep the simulation coherent.

The runtime loop involves Gemini interpreting prompts, planning a sequence of steps, verifying frame success, and adapting to world shifts or user interventions, with Gemini also steering the scene between user inputs to ensure progression.