Google AI Edge Gallery: What It Is and How It Works

Google AI Edge Gallery is Google's own open-source app for running open large language models directly on a phone, tablet, or Mac — no server, no API key, no internet connection once a model is downloaded. It works by wrapping Google's on-device inference stack (LiteRT and LiteRT-LM) in a ready-made interface, so you can chat with a model, ask questions about a photo, transcribe audio, or benchmark performance without writing any code. This guide explains what the app actually does, which platforms and models it supports, how it fits into Google's broader edge-AI tooling, and what its own privacy documentation does and doesn't promise.

What Google AI Edge Gallery actually is

Google AI Edge Gallery is developed and published by Google LLC, and its source code is maintained in the google-ai-edge/gallery repository on GitHub under the Apache License 2.0 — it is genuinely open source, not just free to use. Google's own description, in the app's README, calls it "the premier destination for running the world's most powerful open-source Large Language Models (LLMs) on your mobile device," and positions it as a way to "experience high-performance Generative AI directly on your hardware — fully offline, private, and lightning-fast."

In Google's developer documentation for the underlying LLM Inference API, the app is described more precisely as "an open-source Android application that serves as an interactive playground for developers," built to showcase "the practical implementation of the LLM Inference API and the potential of on-device Generative AI." In other words, Gallery isn't a separate AI model or a chatbot product in its own right — it's the reference app for Google's on-device generative AI tooling, wrapped in a usable interface for anyone, developer or not.

Two of Google's own pages disagree on exactly how mature the project is: the developer documentation labels it "an Alpha release," while the GitHub README's feedback section calls it "an experimental Beta release." Either way, both sources agree the app is under active, fast-moving development, and Google explicitly asks users to report bugs and request features through GitHub Issues rather than treating it as a finished product.

Platforms, requirements, and how to install it

As of the app's September 2, 2026 update (version 1.0.19, per its GitHub release history), Google AI Edge Gallery runs on more than just Android:

• Android: requires Android 12 or later. Install from Google Play, or — for devices without Google Play access — sideload the APK from the project's GitHub Releases page.
• iOS: requires iOS 17.0 or later, available on the App Store.
• macOS: requires macOS 14.0 or later on a Mac with an Apple M1 chip or newer. Google distributes it both as an App Store download and as a standalone DMG file linked from the GitHub README.
• Apple Vision: per the App Store listing, the same app also supports visionOS 1.0 and later.

That multi-platform reach is fairly recent — the project's GitHub wiki still describes iOS as "coming soon" as of this writing, which is out of step with the live App Store listing (Seller: Google LLC) that has required iOS 17.0 for some time. Treat the wiki's platform notes as lagging behind the README and the store listings.

There's no official Windows or Linux build, and no browser-based version despite that being a common search — nothing in Google's GitHub repository, app stores, or developer documentation points to a web app as of September 2026.

Setup itself is short: install the app, then download an open model from inside it (models are fetched over the internet the first time; after that, inference runs without a network connection). The app is a sandbox for the model, not a replacement for one — you still need to pick and download something to talk to.

How it fits into Google's on-device AI stack

Gallery sits on top of a small stack of Google technologies rather than being a standalone product:

• Google AI Edge is the umbrella brand for Google's on-device ML tools and APIs.
• LiteRT (formerly TensorFlow Lite) is the lightweight runtime that actually executes models on-device.
• LiteRT-LM is a newer, LLM-specific layer built on top of LiteRT, and it is now the recommended path for on-device language models.
• MediaPipe is Google's older framework for on-device perception and generative tasks, including the MediaPipe LLM Inference API that Gallery originally showcased.

That last point matters if you're building on top of any of this yourself: Google's developer documentation carries an explicit warning that "the MediaPipe LLM Inference API is in maintenance-only mode," with a direct recommendation to "migrate your Android projects to LiteRT-LM Android (Kotlin) API." Gallery itself has already made that move — its release notes describe importing "LiteRT-LM models using Hugging Face model card URLs," and the model files it uses carry the .litertlm or .task extensions. If you see older tutorials built around the MediaPipe LLM Inference API and the tasks-genai library, know that the API still works but is no longer where Google is putting new development effort.

For a broader look at why running inference locally is a meaningfully different engineering decision than calling a hosted model over an API — in terms of latency, cost, connectivity, and data handling — see our guide to edge AI vs cloud AI. Gallery is a concrete, practical example of the edge side of that trade-off.

Key features

Google's own feature list, consistent across the GitHub README and the Play Store listing, breaks down into a handful of distinct tools rather than a single chat window:

Several of these features depend on the model you've loaded supporting the right modality — a text-only model won't power Ask Image or Audio Scribe. For background on how a single model can handle text, images, and audio together, see our guide to what multimodal AI is.

Which models it supports

Gallery isn't tied to one model. Google's developer documentation describes the app as a way to "discover, download, and experiment with a variety of LiteRT-optimized models from the Hugging Face LiteRT Community and official Google releases." As of the September 2026 update, the headline model is the newly released Gemma 4 family, which Google positions as its most capable on-device model line yet and the one that unlocks Thinking Mode. Google's documentation for the underlying LLM Inference API also points to the multimodal Gemma 3n variants (E2B and E4B) as the models to use for combined image-or-audio-plus-text prompting — a good illustration of how Google's edge and cloud model families relate; we cover that landscape more broadly in our guide to multimodal AI models.

Beyond Google's own releases, the app supports loading custom models: anything already converted to the .litertlm or .task format, including models pulled directly from a Hugging Face model card URL. We could not verify, from Google's own documentation, a specific list of third-party model families (for example, other open-weight LLMs beyond Google's Gemma line) that are officially supported out of the box — treat any such claim you see elsewhere as unconfirmed unless the source names the exact model and format.

What stays on your device — and what Google actually documents

Google's marketing copy for the app is unambiguous: it advertises full offline operation, describing the experience as "fully offline, private, and lightning-fast" and stating that using it means "never sending your data to a server." The project's wiki is more precise about the one exception: the app runs "without needing an internet connection once the model is loaded" — meaning the initial model download does require a network connection, even though the chat, image, and audio processing that follows does not.

On the data-collection side, the app's Play Store data-safety disclosure — provided by the developer, in this case Google itself — states that Gallery "may collect" app-activity and app-performance data, that this data is encrypted in transit, and that no data is shared with third parties. That's a narrower and more concrete claim than "100% private": it means diagnostic and usage telemetry can leave the device, even if your prompts, photos, and audio recordings are processed locally and are not part of what's declared as collected.

Who it's actually useful for

Gallery reads less like a consumer chatbot app and more like a demo environment for a specific audience: developers evaluating whether an on-device model can handle a task before writing integration code, hobbyists who want a private, offline assistant without renting cloud GPU time, and anyone curious about how far phone-class hardware has come for generative AI. Because it's free and open source, it also fits naturally alongside other no-cost options — see our roundup of best free AI tools for comparable software that doesn't require a subscription.

It's worth being clear about what it isn't: it's a software playground for phones, tablets, and laptops, not a hardware product. If your project needs a dedicated accelerator board rather than a general-purpose device — for a robot, a camera rig, or a fixed installation — that's a different category of tool; our guide to best Raspberry Pi AI kits covers that hardware side.

Frequently asked questions about Google AI Edge Gallery

What is Google AI Edge Gallery for?

It's Google's official demo and playground app for running open large language models entirely on-device. It lets you chat with a downloaded model, ask questions about images, transcribe audio, test prompts, and benchmark on-device performance, all without a cloud API.

Is Google AI Edge Gallery free?

Yes. The app is free to download on Google Play and the App Store, and its source code is published under the Apache License 2.0 on GitHub, so it's open source as well as free.

How do you get started with Google AI Edge Gallery?

Confirm your device meets the minimum OS version (Android 12+, iOS 17.0+, or macOS 14.0+ on Apple silicon), install the app from Google Play, the App Store, or the DMG/APK linked from the GitHub repository, then download a model from inside the app. That first download needs an internet connection; using the model afterward does not.

Is Google AI Edge Gallery private?

Prompts, images, and audio are processed on-device rather than sent to a server for inference. However, the app's own Play Store data-safety disclosure states it may collect app-activity and performance data, encrypted in transit, though the developer declares that none of it is shared with third parties.

Is there a version of Google AI Edge Gallery for a computer?

Yes for Mac: Google publishes a native macOS build (macOS 14.0+, Apple M1 or later) through both the App Store and a standalone DMG. There is no official Windows or Linux build, and no evidence of a browser-based version, as of September 2026.

Is the Google AI Edge Gallery source code available?

Yes. The full app is open source under the Apache License 2.0, with the code, issue tracker, and release history published at github.com/google-ai-edge/gallery.