putilov_denis - stock.adobe.com
Unlock the power of local AI with Google AI Edge Gallery
Google AI Edge Gallery lets users run GenAI models locally on Android devices. It supports Gemma models and integrates with Hugging Face.
The beginning of the generative AI boom was largely experienced by users accessing the technology via cloud infrastructure. In that approach, users pay for access to use a given model -- often as a software-as-a-service model.
But there are some drawbacks to running GenAI only in the cloud -- including cost, privacy and potential security concerns. One way to get around those concerns is to run models locally on a user's own device. That's where the Google AI Edge Gallery fits into the GenAI landscape.
The Google AI Edge Gallery was soft-launched in May 2025 as an experimental approach that lets users run GenAI models directly on a mobile device. The initial target is Google Android OS devices, though the plan is to support Apple iOS in the future.
Running models locally means that once the models have been installed on the device, users don't need to be connected to the internet. This allows models to run in remote locations or areas where internet connectivity is not ideal. There is no cost to run the models locally, and it is also ideal for enterprise applications where users can potentially have a higher degree of privacy and security.
How Google AI Edge Gallery works
Instead of relying on cloud resources, Google AI Edge Gallery makes use of a series of technologies that can run locally on user devices.
Google AI Edge Gallery runs as an app on mobile devices. The app is built using the LiteRT -- formerly TensorFlow Lite -- and Media Pipe frameworks that have been optimized for mobile devices. The technology also includes an integrated large language model inference API.
Users can download and run models from the Hugging Face AI model repository that are compatible with LiteRT.
A consideration when using Google AI Edge Gallery is that the AI models operate with fixed knowledge cutoffs determined by their original training data. Unlike cloud-based AI services that can potentially access real-time information or receive frequent updates, the locally downloaded models contain only the information they were trained on before release.
The models do not automatically update themselves in the background, and they don't pull current information from the internet during conversations. Users who want access to more recent information would need to manually download newer versions of models as they become available on Hugging Face.
Installation instructions and system requirements
At least initially, Google AI Edge Gallery is not available via an app store where a simple click would enable installation.
Users need to go through a series of manual steps to install the Edge Gallery app. There are two different ways to install the app -- either directly onto the device or using ADB (Android Debug Bridge). The difference between the two methods is where the app is loaded from. With the direct method, users download and install the app directly on the device. With the ADB method, users download the app to a local computer, then connect the mobile device to that computer and install.
For both the direct and ADB method, the first step is to download the APK file (Android Package Kit file format). This file can be downloaded from GitHub.
Direct mobile installation instructions
- Once the APK file has been downloaded, locate the .apk file in the device's downloads folder.
- Tap on the file to start installation.
- If prompted, enable installation from "unknown sources" in Android's settings. The option is found in Settings > Security/Privacy > Apps.
Instructions for installation via ADB
- Enable developer options on the Android device. Go to Settings > About phone/device, then find the build number entry and click it seven times.
- The next step is to enable USB Debugging. Go to Settings > Developer options and click USB Debugging.
- Connect the Android device to the computer via USB.
- On the computer, open the terminal application or command line and go to the download directory where the APK file is located. Type: run adb install ai-edge-gallery.apk.
System requirements
At launch, all Android devices running the current supported versions should work. Google's documentation does not specifically identify minimum system requirements, though it is clear that larger models require more computing resources and memory.
Supported models
Google AI Edge Gallery supports a variety of models, including the following:
- Gemma. The primary model family is Google's own Gemma 3 open source models. While multiple variants of Gemma 3 are supported, the Gemma 3n models have been specifically optimized for mobile usage.
- Third-party open source models. The app is integrated with Hugging Face, enabling users to browse and download hundreds of open source models.
- LiteRT models. In addition to models found on Hugging Face, Google AI Edge Gallery also supports any local LiteRT task models that developers might have created or have access to run on a mobile device.
Generative AI capabilities
Google AI Edge Gallery has several core capabilities. The three main capabilities that users will see after launching the app include the following:
- AI Chat. This capability provides a natural language interface for a chatbot that is similar to what a user might experience with ChatGPT, except the chatbot runs entirely on the user's device.
- Ask Image. This feature allows users to select an image and ask questions about it. The visual question-answering capability can identify objects, read text from images, solve math problems shown in photos, provide detailed descriptions of scenes, and analyze diagrams and charts.
- Prompt Lab. Unlike AI Chat's conversational format, Prompt Lab focuses on discrete tasks where users input text and receive output. The feature is intended for quick AI tasks, such as content summarization.
Sean Michael Kerner is an IT consultant, technology enthusiast and tinkerer. He has pulled Token Ring, configured NetWare and been known to compile his own Linux kernel. He consults with industry and media organizations on technology issues.
