Google AI Edge Gallery
github.com/google-ai-edge/gallery- Category
- AI Tools
- Rank
- No. 114Tools index
- Pricing
- Open Source
- Type
- APP
- Builder
- google-ai-edge
- GitHub
- 24.7k stars
- Latest release
- 1.0.19
- Date
About
A mobile app that runs powerful open-source AI models like Gemma 4 entirely on your device for completely private, offline AI experiences. Features include AI chat with thinking mode, image analysis, voice transcription, and agent skills that can interact with external tools.
What it does
Google AI Edge Gallery is a hands-on sandbox for comparing local language models on specific hardware. Users can download listed models or import their own, tune generation settings, test prompts, inspect device benchmarks, and extend model behavior through modular skills or MCP servers.
Why it's ranked here
The project combines model evaluation, multimodal tasks, device automation, and extensible tool use in one accessible application. That breadth makes it useful for serious experimentation. Its beta status, demanding model sizes, and experimental MCP support keep it from being a dependable production foundation.
What's good
Prompt Lab exposes temperature and top-k controls instead of hiding model behavior behind presets. Built-in benchmarking measures performance on the actual device. Skills support text instructions, sandboxed JavaScript, interactive views, secrets, and native actions. MCP connections add dynamic tool discovery with per-invocation approval and persistent allow rules.
Tradeoffs
Local models demand substantial storage and memory. Listed downloads range from roughly 555 MB to 4.4 GB, with estimated peak memory between 2.1 GB and 7.0 GB. MCP schemas can consume much of the models’ limited context, smaller models may call tools poorly, and some mobile GPU kernels can distort numerical results. External tools also require network access.
How to use it well
Use it to compare models, prompts, accelerators, and memory demands on hardware you actually plan to support. It suits developers prototyping private mobile features or custom tool flows. Enable only the MCP tools needed for each task. It does not provide offline access to remote services or remove model-specific reliability limits.
Technical notes+
README.md describes Android, iOS, and macOS distribution plus LiteRT and Hugging Face integration. model_allowlist.json records task compatibility, quantization, accelerator choices, download sizes, estimated peak memory, and generation defaults, although its supplied entries cover Gemma 3 and Qwen while README.md announces Gemma 4 support. skills/README.md defines text-only, JavaScript, and native skill paths; JavaScript runs inside a hidden webview through the global ai_edge_gallery_get_result entry point and returns stringified result, error, image, or webview payloads. mcp/README.md documents a StreamableHTTP MCP client with schema injection, permission prompts, custom request headers, and no complete OAuth flow. DEVELOPMENT.md requires a Hugging Face developer application for model downloads in local Android builds. Function_Calling_Guide.md shows Android function tools mapped into app actions.
Observed
- License
- Apache License 2.0
- Platforms
- Android 12 or newer, iOS 17 or newer, and a macOS download
- Install surface
- Google Play, Apple App Store, release APK, and macOS DMG
- Model runtime
- Google AI Edge APIs and LiteRT, with Hugging Face integration for model discovery and download
- Extension interfaces
- Text-only skills, JavaScript webview skills, native app intents, and an experimental MCP client
- MCP transport
- Connects to MCP servers through StreamableHTTP and supports custom HTTP authentication headers
Read from README.md, skills/featured/virtual-piano/scripts/index.js, skills/built-in/calculate-hash/scripts/index.js, skills/featured/restaurant-roulette/scripts/index.js, Android/src/app/src/main/assets/tinygarden/main-K5DSW5YL.js, Android/src/app/src/main/assets/skills/calculate-hash/scripts/index.js, LICENSE, DEVELOPMENT.md, CONTRIBUTING.md, model_allowlist.json, Bug_Reporting_Guide.md, Function_Calling_Guide.md, mcp/README.md, skills/README.md, Android/README.md.
What it can do
Run AI chat conversations with thinking mode
Text prompts and questions → AI-generated responses and conversations
Analyze and describe images
Images or photos → Text descriptions and analysis of image content
Transcribe speech to text
Voice recordings or live audio → Written text transcription
Execute agent skills with external tools
Commands or requests requiring external interactions → Results from external tool integrations and automations
Process AI requests completely offline
Any supported AI task input → AI-generated results without internet connection
Tags
Featured in Intel
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.