THE ESSENTIALS IN 30 SECONDS

  • TechCrunch reports Genie 4 keeps a scene consistent for up to two minutes; Project Genie, built on Genie 3, capped generations at 60 seconds.
  • DeepMind paired Genie 4 with Gemini 4 Argon to simulate a humanoid robot playing checkers, a step toward testing robots in generated worlds.
  • No paper, model card, resolution or benchmark has been published, and Runway's GWM Worlds 2 already streams interactive worlds with audio.

Google DeepMind's Genie 4 launched on 27 October with native audio and live text edits, as covered in our first article. Coverage since then fills in some gaps and shows where the model sits in a fast-moving field.

Longer sessions, more Gemini in the loop

According to TechCrunch, Genie 4 holds a scene together for up to two minutes, which the outlet describes as twice Genie 3's limit. For comparison, Project Genie, the public prototype built on Genie 3, cut generations off at 60 seconds. DeepMind adds that Gemini Omni grounds what the simulation looks like and how it behaves, while a Gemini Flash model plans and corrects in real time. Tejas Kulkarni, a Google research director, told reporters he expects "hybrid world models" in which Gemini writes code and text that Genie turns into a running scene.

From games to robots

The robotics angle is new. DeepMind says it tested Genie 4 with Gemini 4 Argon, which splits an instruction into steps that Genie renders; one demo shows Apptronik's Apollo humanoid playing checkers, entirely in simulation. Another prompt shows a Universal Robots UR5 arm pipetting liquid. The lab links this work to Gemini Robotics. There is precedent: in February, Waymo said its own world model, built on Genie 3, generates rare driving events such as extreme weather or animals on the road.

Open questions

The New Stack points out a basic problem for agent training: when a run fails, it is hard to tell whether the agent or the simulated physics is at fault. Google has not shown Genie 4 can stand in for conventional physics engines, nor published data on whether repeated tests give the same result. DeepMind has released no paper or model card, and its blog gives no resolution, frame rate or latency.

A crowded field

Genie 4 is DeepMind's first world model with built-in audio, but not the only one. In September, Runway previewed GWM Worlds 2, which AlphaSignal reports streams 720p video at 24 frames per second with synchronized audio and no fixed session cap. The same month, AMD agreed to buy Fei-Fei Li's World Labs, maker of the Marble world generator, for $8.2 billion, TechCrunch reported. Independent tests will be needed before anyone can say which approach holds up best.