Gemma Playground: AI Edge Gallery

Ian Ballantyne, Developer Relations Engineer at Google DeepMind, shows what Gemma can do on a Pixel 10 Pro using the Google AI Edge Gallery app. Image understanding, audio transcription, agent skills that route prompts to the right app on your phone, all running locally with no connectivity required.

What’s covered: Using Agent Skills to route a voice prompt to a mood tracker app, taking a photo and outputting structured JSON Schema from it, generating suggestions from an image of plants, recording voice notes for transcription and translation, and identifying objects offline with multimodal image understanding.

Download the Google AI Edge Gallery app on Android or iOS and try Gemma on your own device.

What are you running on-device with Gemma? Drop it in the comments.

Subscribe to Google for Developers → https://goo.gle/developers

Speaker: Ian Ballantyne

Products Mentioned: Google AI, Gemini

Ian Ballantyne demonstrates how Gemma models function locally on mobile devices using the Google AI Edge Gallery app. The showcase highlights practical applications for on-device artificial intelligence, including agent-based task management, image interpretation, and audio processing, all while maintaining functionality in offline environments.

Yes, Gemma is capable of providing home decor advice. Thanks to its multimodal capabilities, it can analyze images of your living space to offer tailored design suggestions.

As showcased in the video, Gemma can process visual input to help with creative arrangements—for example, by suggesting items to add to a desk setup based on a photograph of existing plants (1:43–1:55).

Key ways Gemma can assist with home decor include:

Image Analysis: By taking a photo of a room, Gemma can identify the current layout and provide feedback or redesign ideas.
Creative Recommendations: You can ask specific questions about your space, such as requesting suggestions for furniture, color palettes, or decorative plants to enhance your home.
Local Processing: Because Gemma can run locally on edge devices (like the Pixel 10 Pro), you can generate these design ideas privately and without needing an internet connection.

The Google AI Edge Gallery is an interactive, open-source playground app designed to run lightweight AI models—specifically from the Gemma family—directly on your mobile device (Android and iOS) or laptop. It highlights 100% on-device privacy, requiring no internet connection for inferences. [1, 2, 3, 4]

Core Models: The Gemma 4 Family [1]

The app primarily showcases edge-optimized Gemma 4 models, which turn your device into a powerful local agent: [1, 2]

E2B: Effective 2 billion parameter model.
E4B: Effective 4 billion parameter model.
12B: A larger, more robust model designed for laptops (available on macOS) for advanced reasoning and workflows. [1, 2, 3]

Primary Features

The gallery acts as a sandbox to test autonomous agentic workflows and offline AI capabilities: [1, 2, 3]

Agent Skills: Enables multi-step, autonomous agentic actions, such as reading or creating calendar events and querying external knowledge bases. [1, 2]
AI Chat with Thinking Mode: Allows users to view the model’s step-by-step reasoning process. [1]
Ask Image: Uses your device’s camera or gallery to identify objects, solve visual puzzles, and process multimodal information. [1]
Audio Scribe & Dictation: Transcribes and translates voice recordings and speech in real-time. [1]
Prompt Lab: A dedicated workspace to test specific prompts and control parameters like temperature and top-k. [1, 2]
Mobile Actions & Games: Includes experimental features like offline device controls and interactive mini-games (e.g., Tiny Garden). [1]
Model Management & Benchmarks: Easily download pre-existing open-source models, load your own custom weights, and benchmark device performance. [1]

Where to Get It

Mobile: Available on the Google Play Store and the Apple App Store.
Desktop: You can also run it on macOS laptops to execute agentic scripts on the fly.
Source Code: Available for developers on GitHub. [1, 2, 3, 4, 5]