AGI Pulse
Model releaseGoogle DeepMind · agipulse

Google DeepMind announces Genie 4: real-time interactive world model with native audio, Gemini Omni steering; research preview then Ultra/Project Genie + API early access

8 reads

Published
2026-10-28
Type
Model release
Organization
Google DeepMind
Source
agipulse

Why it matters

【Key points】

Google DeepMind on 2026-10-27 (~16:00–16:13 UTC) announced Genie 4, a real-time interactive world model. Primaries: https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ (article:published_time 2026-10-27T16:13:00Z) and https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/. Site has no prior Genie event (max e/1060).

From the primaries:

1) Real-time interactive world sim; claims higher realism vs Genie 3; first DeepMind real-time world model with native audio (ambience, music, dialogue).

2) Gemini Omni anchors appearance/behavior; other Gemini keep coherence; natural-language steering plus keyboard navigation.

3) Physics and people claims shown in vendor demos.

4) Research preview for trusted testers; planned for Google AI Ultra / Project Genie and early-access Gemini API / AI Studio partners; public waitlist. Not full GA; no public prices or independent leaderboard.

【Verifiable】

DeepMind post; Google blog mirror; no site Genie event; deepmind_rss / gblog_* STRICT-hit same URL cluster.

【Our read】

1. Background: world models move from generate-a-clip toward playable, audible, steerable sims. Genie 3 (Aug 2025) set DeepMind’s line; Genie 4 wires Gemini Omni + reasoning with native audio into Project Genie / Ultra / API.

2. Peers: offline video world models (e.g. World Labs Marble); explicit-3D tools (e.g. Viggle JST-2); embodied/game agents (SIMA, Gemini 4 Argon) that act in worlds rather than generate them. Genie 4 is real-time multimodal sim + language-steer. No public numbers—do not call a clean win over Marble/JST-2.

3. Impact: if Ultra / Project Genie and API early access land, creators and robot/agent training get steerable scenes with dialogue. Still gated by preview/waitlist near term.

4. Limits: occurred_at ≈ 2026-10-27T16:00:00Z (meta 16:13Z). Demos self-reported. GO: full primary, no duplicate; same-round MAI image bumps, Argon marketing, Qwen Code minors → SKIP.

Read original →

Related developments

Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy