Genie 3 Unveiled: DeepMind's General World Model Generates Explorable Worlds From a Sentence

Google DeepMind announced Genie 3 on August 5, 2025 — a general-purpose world model: give it a text prompt and it generates a dynamic world you can navigate in real time. The official spec is 24 frames per second at 720p, holding consistency for a few minutes (DeepMind blog, 2025-08-05).

The first world model with real-time interaction#

The blog's wording: Genie 3 is DeepMind's first world model to allow interaction in real time, while improving consistency and realism over Genie 2. The capabilities it lists are mostly environmental — for example natural phenomena such as water and lighting, and complex environmental interactions.

The limits sit on the same page, and are more specific than the capability list:

• Limited action space: promptable world events are not necessarily performed by the agent itself, and the range of actions an agent can perform directly is constrained.
• Multi-agent interaction: accurately modelling complex interactions between multiple independent agents in shared environments remains an open research challenge.
• Real-world locations: it cannot simulate real-world locations with perfect geographic accuracy.
• Text rendering: clear, legible text is often generated only when it is provided in the input world description.
• Interaction duration: currently a few minutes of continuous interaction, not extended hours.

Why DeepMind puts it next to AGI#

The blog's own words are that world models are "a key stepping stone on the path to AGI," because they make an unlimited curriculum of simulated environments available to agents. To test that, DeepMind generated worlds for SIMA, its generalist agent for 3D virtual settings — the agent is given goals in a generated world and sends navigation actions to Genie 3, which is unaware of the goal and simply simulates the future from the agent's actions. The blog says better consistency is what makes longer action sequences, and more complex goals, possible.

Project Genie: how to try it now#

The route to hands-on access is Google Labs' Project Genie, which the official page calls an "early research prototype": you can create worlds from text, images or even Street View imagery and explore them by walking, riding, flying or driving; access requires a Google AI Ultra subscription (Project Genie official page).

(Revised October 12, 2026: the TechCrunch attribution was replaced with DeepMind's own blog wording that world models are "a key stepping stone on the path to AGI"; the unverifiable claim that DeepMind opened public access via Project Genie by year's end was removed and replaced with the official page's "early research prototype" description and subscription requirement.)