Updated on August 17, 2026
Google AI Edge is Google's development stack for executing AI models directly on a device, with no server round-trip at all. Everything ships free and open source, from the LiteRT engine to the AI Edge Gallery demo app. You convert a PyTorch, JAX, Keras or TensorFlow model once, then it executes with CPU, GPU or NPU acceleration across Android, iOS, the web and embedded boards.
LiteRT is the central engine of the stack and the official successor to TensorFlow Lite. Its runtime weighs only a few megabytes and accelerates converted models on CPU, GPU and NPU. On top of it, LiteRT-LM orchestrates large language models and executes the same LLM on Android, iOS and the web without a rewrite.
MediaPipe adds ready-made APIs (background blur, object detection, hand tracking), while Model Explorer maps out a model's architecture to spot bottlenecks. The whole approach is edge computing in practice, moving the compute to the data rather than the other way around.
| Component | Role in the stack |
|---|---|
| LiteRT | Model conversion and accelerated execution |
| LiteRT-LM | The same LLM on Android, iOS, web and embedded |
| MediaPipe | Low-code APIs for vision, text and audio |
| Model Explorer | Architecture visualization and debugging |
| AI Edge Portal | Large-scale benchmarking on real phones |
AI Edge Gallery, available on Android and iOS, lets you test open models like Gemma 4 without writing a line of code. You pull a model from Hugging Face, switch to airplane mode, and the conversation keeps going, pictures and audio included (models weigh several gigabytes, so your data plan will thank you for using wifi).
Here is what the app packs in, feature by feature.
The code, written in Kotlin on Android and Swift on iOS, doubles as a reference for wiring LiteRT-LM into your own apps, in the same spirit as other major open source AI projects hosted on GitHub.
AI Edge Portal, the Google Cloud side of the ecosystem, puts your models through a physical fleet of more than 120 Android devices. Reports break down initialization time, generation speed and peak memory, so you know your app will hold up on your customers' phones before release.
Beyond that, getting started costs nothing. Libraries install from GitHub, pre-converted models sit on the LiteRT Hugging Face community, and the Gallery app downloads from the usual app stores.
Yes, the entire stack is free and open source, with LiteRT and the AI Edge Gallery app published under the Apache 2.0 license. Open models like Gemma download at no cost from Hugging Face. Only AI Edge Portal sits within Google Cloud, under that platform's own terms.
On a phone, yes, that is precisely its home turf. Ollama targets computers (Mac, Windows, Linux), while AI Edge Gallery executes Gemma and other open models on Android and iOS, fully offline. The two complement each other rather than compete, depending on which device you have in hand.
Android 12 or iOS 17 at minimum, according to the project's official requirements. In practice, larger models perform far better on memory-rich hardware such as recent Pixel or iPhone flagships, while mid-range phones should stick to the compact models in the catalog.
LiteRT-LM, the stack's LLM layer, is the same infrastructure Google itself relies on to deploy Gemini Nano in Chrome and on the Pixel Watch. You get the engine already proven in Google's own production systems to embed your own language models into your apps.
Verdict: No inference bill, no data leaving the device, and a marked path from conversion to deployment. Mobile and web developers get the full toolchain here, while the simply curious can size it all up in minutes with the Gallery app.
GPT-5.3-Codex 
Claude Code 
Cursor AI 
