Skip to content
RRanked By Task

This Week

DeepMind previews Genie 4, a steerable world model

Genie 4 is a research preview for trusted testers. It simulates a scene in real time, with native audio, and is not a general video tool yet.

Official source (opens in a new tab)

Google DeepMind's post on 27 October 2026 announces Genie 4, a world model for real-time interactive simulation. DeepMind says it is natively multimodal, so music-like audio, ambient noise, and interactive dialogue are generated with the world. Gemini Omni anchors how a simulation looks and behaves. Other Gemini models help keep it coherent. You can start an event, and Genie 4 simulates what happens next, including how the people in it move, react, and speak.

The post says you steer a running simulation with natural language, such as changing the weather, moving the camera, or adding a person or object. Keyboard controls handle navigation and the camera. Between those inputs, Gemini monitors the scene. Gemini 3.8 Flash is the guiding agent in the loop: it watches frames, breaks a goal into steps, and adapts the plan when the world shifts or you intervene. DeepMind also describes experiments that pair Gemini 4 Argon with Genie 4, including a simulation of Apptronik's Apollo humanoid playing checkers. The post calls Genie 4 DeepMind's first real-time world model to generate audio natively. Video and audio from Genie 4 include SynthID watermarks.

Genie 4 is a research preview for trusted testers. DeepMind says a broader release is coming, starting with Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio. The post says the model is still early. Visual details and object permanence can drift in a longer session, after several complex prompt edits, or while moving through a scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency are listed as ongoing research.

Who it's for: a trusted tester, or someone waiting on the Ultra and API rollout, who wants a steerable simulation rather than a finished shot.

Fits: Best for Video stays Runway, then Kling, then Luma, then Google Veo, then Pika. Genie 4 is not a clip you open in an editor, and it is not generally available. The video guide is still about shots you might cut. The verdict is unchanged.