Gemini’s Multimodal Model Lineup

▶ Watch (02:18)

Google DeepMind shipped more than a dozen models in six weeks. Gemini 3.1 Pro, Flash, and Flash Light handle most production use cases. The embeddings model supports video, images, audio, text, and code in one shared space. Nano Banana 2 handles image generation. LIA 3 handles music. VO3.1 Light handles video at a low cost profile. Gemma 4, the open model family, runs on mobile devices and carries an Apache 2 license.

AI Studio: Video Analysis and URL Grounding

▶ Watch (08:40)

AI Studio accepts YouTube URLs directly as inference input. Five minutes of video runs to about 27,600 tokens. With grounding enabled, the dinosaur video analysis demo (09:46) shows the model creating a timestamped dinosaur table and correctly noting pteranodons are pterosaurs, not dinosaurs. URL context lets the model pull in external pages and cite each source inline. A “get code” button outputs the Python, TypeScript, or Java needed to replicate any experiment from the UI.

Code Execution and Compare Mode

▶ Watch (13:16)

Code execution gives Gemini a sandboxed Python environment with data science libraries pre-installed. It runs as a tool call with no impact on the local environment. In the Lego bricks bounding box in compare mode (14:20), Gemini 3.1 Flash Light identified every green brick and drew bounding boxes in one pass. The total cost came in at well under a fraction of a penny. Gemini 3 Flash reached the same answer, just slightly slower.

Gemini Live: Real-Time Screen and Voice Interaction

▶ Watch (19:12)

Gemini Live streams screen output, camera video, or audio and responds in real time. It supports function calling and Google Search grounding inside a live session. In the Gemini Live screen-sharing demo (20:26), the model described a browser window, switched to Italian on request, then delivered a poem in a Texan accent with no configuration change between turns.

“Well, howdy there. Let me take a gander” — Gemini Live (model response)

Real applications include bank kiosk multilingual support, AR navigation, and orchestrating the Pupper open-source robot.

AI Studio Build and the Generative Media Stack

▶ Watch (25:04)

AI Studio Build generates full-stack apps from a voice prompt, with Firebase database and Google OAuth bundled in. In the bookshelf cataloger built live (26:47), Gemini 3.1 Pro built, deployed, and QR-coded a working app that identifies book titles from a shelf photo, enriches them with Google Search, and persists them per user. One-click deploy to Cloud Run ships it publicly. Alongside Build, VO3.1 Light generates 720p video with audio at low cost, LIA 3 generates full songs with lyrics in any language, and Nano Banana 2 handles image editing with reverse image search grounding.

Q&A

Is there a plan for a standalone AI Studio app? Bailey confirmed there is a plan for an AI Studio app, citing Logan’s hints on Twitter, and described the team as excited. ▶ 1:01:44

Notable Quotes

Well, howdy there. Let me take a gander Gemini Live (model response) · ▶ 22:22

automatically catalogs all of your books Paige Bailey · ▶ 28:20

of models. So it’s not just one model. Paige Bailey · ▶ 33:43

Key Takeaways

  • Gemini 3.1 Flash Light processes a YouTube video for under a penny per request.
  • AI Studio Build creates full-stack apps with Firebase and Google Auth from a voice prompt.
  • Genie 3 generates navigable game worlds frame by frame with no traditional physics engine.