Google AI Edge Gallery: run open LLMs on your phone, no cloud, no bill
Updated on August 13, 2026
Google AI Edge Gallery is a free, open-source app from Google that runs generative AI models directly on your smartphone, with nothing sent to a server. After a one-time model download, everything works offline, from multimodal chat to audio transcription. The app is available on both Android and iOS, still labeled experimental, and released under the Apache 2.0 license with no subscription attached.
- Completely free, open source, ad-free
- Works offline, data stays on the device
- Multimodal chat with images and audio
- Built-in benchmark for your own hardware
- Available on Android and iOS
- Still an experimental app, occasional crashes reported
- Larger models call for a powerful phone
- English-only interface for now
Google AI Edge Gallery, a local playground for Gemma
Google AI Edge Gallery downloads open models such as Gemma 4 and FunctionGemma from Hugging Face, then runs them on your phone's CPU or GPU. Nothing travels through the cloud, which settles the privacy question for prompts, photos and voice recordings from the start.
The app first spread as an experimental APK, racking up 500,000 downloads in two months before landing on the Play Store and later the App Store. Its full source code sits on GitHub alongside other open-source AI projects maintained by Google.
Six workspaces, from multimodal chat to a mini-game
Every use case gets its own tile on the home screen. You tap Ask Image, snap a photo of a dubious mushroom on a hike, and the model answers where your data plan gives up (a scenario that keeps popping up in reviews, and it holds up). Thinking Mode, limited to the Gemma 4 family, even displays the model's step-by-step reasoning.
| Workspace | What you do there |
|---|---|
| AI Chat | Multi-turn conversations with images and audio |
| Ask Image | Questions about a photo or the camera feed |
| Audio Scribe | Voice transcription and translation |
| Prompt Lab | Single-turn prompt tests with temperature and top-k controls |
| Mobile Actions | Flashlight, volume and apps driven by plain language |
| Tiny Garden | A virtual garden you tend by talking |
Getting started and squeezing out performance
The app installs free from the Play Store and the App Store, with no waitlist and no developer account. Heavier models do ask for serious hardware, and 8 GB of RAM makes a real difference.
To find out what your phone is worth, a built-in benchmark measures each model's speed on your chip. Chat history restores almost instantly, with prefill speeds above 3,000 tokens per second on recent GPUs. Tinkerers can even hook up external MCP servers, an experimental option on Android.
Frequently asked questions
Is Google AI Edge Gallery free?
Yes, the app is entirely free, with no subscription or in-app purchases. The project is open source under the Apache 2.0 license, so the code can be reused, including commercially. The only real cost is the storage space the downloaded models take up on your phone.
Does Google AI Edge Gallery work without internet?
Yes, once a model is installed, every answer is computed on the device, airplane mode included. A connection is only required to download models in the first place, or to reach optional external tools such as MCP servers if you switch those on.
Do I have to create a Hugging Face account?
Not to install the app, but downloading some models goes through Hugging Face and requires signing in to accept their license. Creating the account is free and takes a couple of minutes, something worth doing before heading somewhere with no signal.
Can I load my own models into Google AI Edge Gallery?
Yes, custom models are supported on top of the built-in list. They must be converted to the LiteRT format used by Google's runtime, though, since standard.safetensors files will not load directly, a limitation frequently raised on forums and in app reviews.
Verdict: Zero subscription, zero server, zero data leaving the device. Mobile developers and anyone curious about local AI get a ready-made proving ground to see what a phone can run on its own.
