Overview
Google AI Edge Gallery is an app for trying open models on the device in your hand instead of a server. Every inference runs on the local hardware, so prompts, images and other data never leave the phone and the whole thing works with no network connection. It is built on Google AI Edge and the LiteRT runtime, with Hugging Face integration for discovering and downloading models.
The app is organised as a set of tiles rather than a single chat window. AI Chat carries multi-turn conversations and a Thinking Mode that exposes the model's step-by-step reasoning on supported models, starting with the Gemma 4 family. Ask Image answers questions about a photo or camera frame. Audio Scribe transcribes and translates recordings. Prompt Lab is a single-turn workspace with direct control over parameters like temperature and top-k. Agent Skills augments a model with tools — Wikipedia grounding, interactive maps, visual summary cards — and can load modular skills from a URL.
It doubles as an evaluation harness. Model management lets you download from the built-in list or side-load your own model, and the benchmark tool reports how each one actually performs on your specific hardware — which is the number that matters when the constraint is a phone rather than a datacenter. Two smaller tiles, Mobile Actions and Tiny Garden, are driven by a fine-tune of FunctionGemma 270m. The project describes itself as an experimental beta.
What it does
- 100% on-device inference — no internet required, and prompts, images and files never leave the hardware
- Runs the Gemma 4 family with Thinking Mode, which shows the model's step-by-step reasoning on supported models
- Ask Image for multimodal questions from the camera or gallery, and Audio Scribe for on-device transcription and translation
- Agent Skills tile that adds tools such as Wikipedia grounding, interactive maps and visual summary cards, loadable from a URL
- Prompt Lab for single-turn experiments with granular control over temperature, top-k and other parameters
- Model management plus a benchmark tool that measures each model on your own device
- Side-load your own models alongside the curated download list, via Hugging Face and LiteRT
- Ships for Android 12+, iOS 17+ and macOS, with an APK for devices without Google Play
Getting started
Gallery is distributed as an app, not a library. It needs Android 12 or newer, or iOS 17 or newer; a macOS build is also published.
Install the app
Get it from Google Play or the App Store, download the macOS DMG, or — on devices without Google Play — install the APK from the latest GitHub release. The project wiki has detailed instructions, including for corporate-managed devices.
Download a model
Open model management and pull a model from the built-in list. Models are fetched through the Hugging Face integration and stored on the device for offline use.
Pick a tile
AI Chat for conversation with optional Thinking Mode, Ask Image for questions about a photo, Audio Scribe for transcription and translation, or Prompt Lab for single-turn prompts with parameter control.
Benchmark on your own hardware
Run the benchmark tool against each downloaded model to see what it does on your specific device, which is the figure that decides whether a model is usable on that phone.
Build it yourself
The repository carries development notes for building the app locally.
git clone https://github.com/google-ai-edge/gallery.gitCommands and code are distilled from the project's own documentation — always check the official repo for the latest.
When to use it
- Reach for it to try an open model on a phone before committing to shipping it in your own mobile app
- Reach for it when a model has to answer with no network connection at all
- Reach for it to benchmark candidate on-device models against the hardware you actually target
- Reach for it to see how agent skills and tool use behave inside a small on-device model
How Google AI Edge Gallery compares
Google AI Edge Gallery alongside other open-source local runtimes tools AI/TLDR tracks, ranked by GitHub stars.
| Tool | Stars | What it does |
|---|---|---|
| Ollama | ★ 181k | A developer-friendly tool that downloads and runs local LLMs from the terminal with a built-in OpenAI-compatible API. |
| llama.cpp | ★ 128k | A C/C++ inference engine that runs LLMs in the GGUF format on CPUs, Apple Silicon, and GPUs with low memory use. |
| GPT4All | ★ 77.4k | GPT4All is a free desktop app and Python client that runs large language models locally on your own computer, with no API calls or GPU required. |
| LocalAI | ★ 49k | A self-hosted server that exposes an OpenAI-compatible API for running text, vision, voice, and image models on local hardware. |
| Jan | ★ 44.4k | An open-source desktop app that runs LLMs fully offline as a ChatGPT-style assistant on your own computer. |
| AirLLM | ★ 34k | A Python inference library that keeps only one transformer layer on the GPU at a time, so a 70B model runs on a single 4GB card and a 671B MoE model on about 12GB, without quantization. |
| Colibrì | ★ 27.3k | A pure-C inference engine that keeps a Mixture-of-Experts model's dense trunk resident in RAM and streams its routed experts from disk, so 744B-2.8T models run on consumer hardware. |
| Google AI Edge Gallery | — | Run open models entirely on your phone, offline, and benchmark them there |