In an increasingly interconnected world, the ability to access advanced artificial intelligence (AI) tools without relying on the cloud has become a priority.
Google AI Edge Gallery (GAEG) arises as a response to this need: an application designed for users to run AI models directly on their mobile devices. **
Discreetly presented last Saturday, this smart tool from Gemini opens the door to an environment where privacy, speed and accessibility do not depend exclusively on an Internet connection.
Let’s explore in detail its features, its limitations, its target audience, and how it compares to other solutions on the market.
Does it make sense to run AI models on mobile?
The project was announced discreetly, without major announcements and without being available in the Google Play Store, since it is distributed for now as an APK file from the Google AI Edge GitHub.
Its arrival occurred in a context in which more and more users demand to be able to use generative AI functions (such as creating images or writing texts) on their smartphones without resorting to remote servers.
The reasons go beyond simple technological curiosity: there are regions with reduced connectivity, travelers who do not always have data available or professionals who handle sensitive information and prefer to keep their data locally.
Google Vision
With GAEG, Google aims to decentralize part of the power of its AI models. The company has been working on optimizing algorithms so that they adapt to the architecture of mobile processors (GPU and NPU) and can run fluently on Android devices.
The ultimate goal is to offer an experience where AI tasks, such as answering complex questions or generating code snippets, occur directly on the phone, without the need for a cloud backend.
What can Google AI Edge Gallery do?
GAEG incorporates a series of functionalities that today we associate with cutting-edge AI models.
These include the generation of images from textual descriptions, the answer to text queries or images uploaded by the user and the possibility of writing or editing code conversationally.
Image Generation and Analysis
One of GAEG’s great bets is the creation of illustrations, diagrams or artistic compositions based on a short text.
This capability is supported by AI models developed by Google (such as Gemma 3n) and others available in external repositories (for example, Hugging Face), which the user can choose based on the size and complexity of the model.
When running locally, response time may vary depending on hardware power, but the main advantage is that no generated image or sensitive data leaves the device.
Text and Code Tasks
The text processing function is not limited only to answering questions: GAEG can summarize documents, rewrite paragraphs or generate code fragments in different languages.
For developers, this means having a wizard on hand that, in your pocket, can help from debugging errors to generating usage examples. All this without sharing the content with external servers, which increases privacy and speeds up execution in certain scenarios.
Multi-shift Conversations
Beyond specific responses, GAEG supports multi-turn dialogues. This allows for more natural conversations with the AI model, whether to resolve questions about a project, explore learning topics, or even practice languages.
The app offers real-time metrics (such as response time to first token or decoding speed) so the user can better understand the performance of each model.
##How does Google AI Edge Gallery work inside?
The key piece that enables AI execution on mobile phones is LiteRT (Lite Runtime), an optimized engine capable of managing model inference on local hardware.
LiteRT takes care of converting the model operations to a format suitable for the available accelerators (GPU, NPU or even the CPU), minimizing latency and maximizing energy efficiency.
Thanks to this adaptation, GAEG can significantly reduce dependence on the cloud.
Privacy and Security
One of the strongest arguments in favor of GAEG is that all the data it processes remains on the device. This prevents personal or sensitive information from circulating on the Internet, reducing the risk of leaks.
Additionally, since it is open source software under the Apache 2.0 license, the community can audit the code and ensure that there are no hidden elements that compromise privacy.
Technical Requirements
To install GAEG, the user must download the APK from the official repository on GitHub.
It is necessary to have an Android device whose specifications meet at least a certain level of storage (as models range between 560 MB and 4.4 GB) and a version of Android compatible with the LiteRT libraries.
Due to its experimental status (alpha stage), not all mobile phones on the market guarantee optimal performance, and some may experience delays or cuts in complex tasks.
Is Google AI Edge Gallery worth installing?
Google AI Edge Gallery is currently in an early phase where manual installation from GitHub and the still unpolished interface limit its adoption to technical users with high-end devices.
Compared to solutions such as Ollama, GPT4All, LM Studio or Phi3 mini, GAEG stands out for its native integration into Android and the LiteRT engine, which automatically adapts the models to mobile hardware.
This advantage gives it unique potential to run AI on the phone faster and more privately than in competing desktop environments.
Looking ahead, the addition to the Google Play Store and the launch of the iOS version will simplify installation and expand the target audience.
As developers contribute more models adapted to LiteRT and Google debugs this alpha version, the catalog of possibilities will grow.
If the project overcomes current performance and usability barriers, GAEG could become a decisive push for hundreds of thousands of users to work with advanced AI without relying on the cloud, fostering innovation and protecting privacy even in environments without a reliable connection.
Coughing up millions of parameters is impractical, since inference times skyrocket and end up consuming a considerable part of the RAM memory.
This post is also available in: