Omni

Confidence 0.70 · 1 source · last confirmed 2026-06-15

Google DeepMind’s single any-input → any-output multimodal model, launched at Google I/O 2026 and described by Logan Kilpatrick in the Sequoia interview. Part of Google.

What it is

• One model, not routing. A true omni model that replaces ~8 previously separate systems Google trained — text (baseline Gemini), audio, music (Lyria), image (nano banana), and video (Veo). Kilpatrick stresses it is “not routing to a bunch of different models.”
• A “world model” by the blurring definition. Demis Hassabis framed Omni as a world model for its level of world understanding; Kilpatrick notes the term world model is blurring from “action-conditioned video model” toward “a model with understanding of the world.” An instance of the foundation-models trend toward consolidation.
• First iteration = Omni Flash, currently strongest at video editing (the real-time stage-edit “dog” demo); other modalities functional but not yet state-of-the-art.

A near-term implication on the wiki’s radar: vibe-coded games will be coding-agent + game-engine based, not pure world-model-based, until world models gain the scaffolding to be grounded in reusable use cases.