Google for Developers introduced Genie 4 on Release Notes, with host Logan Kilpatrick joined by researchers Raia, Diego and Hang to demo the new world model.
The update is the first Genie release to ship through products and an API, Kilpatrick said.
A world model is a system that keeps an internal state of a grounded environment and predicts how that state changes when you act, so it can simulate collisions, movement and other physics even in scenes it has never seen.
DeepMind started the Genie line to generate endless game-like simulations to train agents, Hang said, and Genie 1 was a high-risk 2D experiment that proved a model could learn to move left, right and jump from platformer video. Genie 2 moved that idea to 3D but still looked like games and was not real-time, and Genie 3 in January brought richer 3D realism and real-time interaction for the first time, the team said.
Genie 4 pushes three things at once, and the demos were set up to compare it directly to Genie 3. Diego showed a Genie 3 starting image that produced a short silent clip with gamified characters, then switched to Genie 4 on the same prompt where a band played with native sound, and later an egocentric view of hands making coffee. He noted the model predicts video and audio together from one model, not as a dub added after.
That realism comes in part from rebasing on Gemini Omni Flash, Hang said. Real-world video is always paired with in-sync audio, so training on multimodal data lets the model treat sight and sound as two components of the same world, rather than two separate tasks. DeepMind adds that Genie 4 also brings advanced physics and natural human motion, which the Release Notes demo illustrated with close camera moves and subtle hand motion that were softer in the prior gamified look.
Interaction also changed. Genie 3 centered on third-person characters like a dragon toy you could walk around, Diego said, while Genie 4 emphasizes first-person navigation with keyboard and mouse and direct camera steering you saw in the band and coffee examples. For research, the demo added pause and resume so you can type text prompts mid-stream without the world running away from you, and the team said it is still testing whether queuing prompts, interleaving captions live, or adding procedural actions is the right way to steer.
Dialogue is the other new audio layer. In the coffee shop clip, Diego dropped the cup and a character replied, "No worries at all. Accidents happen." The exchange was not just Genie 4 lip-syncing, Hang said. A Gemini model in the loop watches the last frames, infers user intent, and drafts what should happen next, including speech. The live demo used vanilla Gemini 3.8 Flash to keep latency low, the team said, while heavier experiments use Argon.
That Gemini harness matters for two applications the team called out, education and robotics. For instruction, Gemini brings world knowledge and reasoning to decide how to respond to your action. For robotics, seeing tasks from a first-person view, including how a heavy cup versus a light cup behaves or what happens when you spill, could help robots predict short-term physics and move more like humans, Raia said. A recorded robotics demo showed the pattern: Argon wrote step-by-step captions to arrange Scrabble tiles to spell a word, and Genie 4 rendered the simulation and the interleaved arm actions from those instructions, though that orchestration was not running in real time.
Genie 4 is still research, the researchers stressed. It will make mistakes, dialogue can artifact, and a cohesive conversation that stays grounded in the scene remains hard when everything must run in real time.
One open technical question lingered at the top of the show. When Kilpatrick joked about giving all the TPUs to one user, he asked if it could run at 60 FPS, and Hang said they had never tried and did not know.
