Google DeepMind announces Genie 4 — most realistic and interactive world model yet

TL;DR

Google DeepMind announces Genie 4, framed as their most realistic and interactive world model yet. The launch highlights four claimed capabilities: photorealism built on Gemini Omni, advanced physics (aerodynamics, collisions, naturally-moving people), real-time controls (prompt-driven scene changes and interaction with world elements), and multimodal output including synchronized sound, speech, and reactive characters. The tweet is a product-marketing announcement with embedded demo videos — no architecture, parameter count, benchmark scores, availability tier, or technical paper is disclosed.

Key claims

Genie 4 is positioned as the newest entry in DeepMind’s interactive-WFM product line following Genie 3 [tweet body].
Built on Gemini Omni as the underlying backbone for photorealistic lifelike detail across spaces and subjects [thread post 2].
Advertises aerodynamics, collisions, and natural human motion as a physics-tier claim [thread post 2].
Supports real-time prompt-driven scene changes, element addition, and interaction with world content during exploration [thread post 2].
Characters inside the world can move, react, and speak — bundling audio-video generation into the interactive rollout [tweet body].

Method

Not disclosed. The announcement reveals only that Genie 4 is “built on Gemini Omni” as the backbone, with no information about parameter count, training data, architecture, context length, resolution, framerate, rollout duration cap, or availability. Demo videos are embedded but opaque to text extraction.

Results

No benchmarks, head-to-head comparisons against prior Genie versions, Cosmos 3, LingBot-World, or any evaluation metrics are disclosed in the launch tweet. The only quantitative signals are the embedded demo videos (2:23 main video + four shorter clips on the follow-up post).

Why it’s interesting

Genie 4 extends the closed-flagship interactive-WFM product line DeepMind established with Genie 3: A new frontier for world models Genie 3 and productized as Project Genie: Experimenting with infinite, interactive worlds Project Genie — now with the explicit claim of integrated audio (sound + speech) and character reactivity, which were not foregrounded in Genie 3. The “built on Gemini Omni” phrasing is new: it reframes Genie 4 as a specialization of the Gemini Omni backbone (see Introducing Gemini Omni), paralleling the pattern Waymo used to post-train Genie 3 into a camera+lidar driving simulator — a reuse pattern World Foundation Models has been tracking as “closed flagships specialize into vertical foundation models via post-training.” It also lands against a crowded 2026 open/closed WFM field where Advancing Open-source World Models (LingBot-World) LingBot-World, Meet Odyssey-3: Our Most Powerful Foundation World Model Odyssey-3, and Cosmos 3: Omnimodal World Models for Physical AI Cosmos 3 have each tried to narrow the gap to closed flagships; the sibling of this launch, Google DeepMind announces Gemini 4 Argon — new frontier model with 1M-token output for coding, enterprise knowledge work, and cybersecurity defense Gemini 4 Argon, hints that DeepMind is coordinating the WFM and reasoning tiers simultaneously.