Genie 3

Google DeepMind, United Kingdom · August 2025

World model · Closed weights

Parameters
not published
Not published.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

Genie 3 is Google DeepMind's general purpose world model: from a text prompt it generates an interactive environment that can be navigated in real time at 24 frames per second and 720p, staying consistent for a few minutes with a visual memory that reaches back about one minute. Promptable world events can change the weather or add objects and characters while the world runs. It is not robot specific; DeepMind tested it with its SIMA agent, which sent navigation actions while pursuing goals the world model did not know. It is a limited research preview with no public weights and no published parameter count.

Architecture

Backbone
Not disclosed.
Action head
No action output. It takes navigation actions from a user or an agent and generates the next frames.
Parameters
Not published.
Pretraining data
Not disclosed.
Embodiments
Virtual agents (SIMA) in generated worlds; no physical robot reported

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Not published. Access is limited to a small group of academics and creators.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

None reported. The robotics relevance is indirect: training and testing embodied agents in generated worlds, shown with the SIMA agent.

Fine tuning it yourself

Not possible. There are no public weights, and access is a limited research preview.

Where it helps, where it does not

Strengths

  • Real-time interaction at 24 frames per second and 720p, consistent for a few minutes.
  • World events can be changed by prompt while an agent acts, which makes it possible to test agents under changing conditions.
  • Shows that one learned model can serve as the environment for an embodied agent.

Limits

  • Closed: limited research preview, no weights, no published size.
  • Limited action space; agents cannot perform the full range of actions.
  • Interactions between several independent agents are still a challenge, according to DeepMind.
  • Sessions last minutes, not hours.
  • No robot manipulation results.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-12.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.