Gemini Omni demos: 9 Google I/O videos show multimodal power

Google published nine demo videos after Google I/O 2026 that illustrate how Gemini Omni and Gemini 3.5 handle multimodal tasks and conversational workloads. The clips, posted to the company blog and official channels, present discrete use cases ranging from image and audio understanding to code generation and agent-style task execution.

What the videos show

The set of videos separates into two rough groups. Gemini Omni clips focus on multimodal interaction, combining images, video, and audio with conversational prompts. Examples include identifying objects in cluttered scenes, following a narrated walkthrough of a physical space, and answering questions about video content. One clip shows real-time annotation of a scene, with the model responding to both a live camera feed and spoken instructions.

Gemini 3.5 demos emphasize text and code capabilities. Several videos show stepwise problem solving in a chat interface, including debugging and generating code snippets, producing structured outputs, and integrating external tools for retrieval and computation. One example demonstrates a multi-step planning task where the model drafts a sequence of actions, queries a simulated calendar, and refines the plan after receiving new constraints.

Across the videos the interface variety is notable. Some demos run inside a chat console, others overlay model outputs on top of video playback or AR-style visuals. Inputs are often multimodal: screenshots or phone-camera footage plus spoken queries, or a developer prompt paired with a codebase snapshot. Each clip highlights a single capability, keeping interactions short and focused.

Technical highlights and limits

The demos illustrate two technical themes: broader input modality support and tighter tool integration. Gemini Omni is shown accepting images, short video clips, and audio cues within a single session, and combining those inputs to produce grounded answers. Gemini 3.5 is presented as an updated chat-and-code workhorse, with extended context handling for multi-step tasks and explicit tool calls for retrieval or execution.

The videos do not disclose behind-the-scenes deployment details, latency figures, or the compute used for each demo. Interactions appear curated and often rely on clean inputs: well-lit video, clear audio, and orderly code samples. That curation makes capabilities easier to demonstrate but limits evidence about robustness in noisy, real-world conditions. The clips also show built-in guardrails in some cases, for example model refusals on sensitive prompts, but do not provide a systematic assessment of failure modes.

Here is a concise comparison of the nine demos and what each illustrates:

Why it matters

The videos make clear that Google is positioning multimodal models as practical interfaces for mixed visual, audio, and text tasks, while keeping a separate line for text-and-code-intensive assistant work. Developers and enterprises preparing to integrate multimodal inputs will need to evaluate robustness and privacy in real deployments, since the demos show capabilities but not real-world failure rates or infrastructure costs.