Google's pocket-sized LLM showroom
A mobile app that lets you download, benchmark, and actually run open-source generative AI models entirely offline.

What it does
AI Edge Gallery is a Kotlin-based Android/iOS app that functions as a sandbox for on-device generative AI. You browse a model catalog (pulled from Hugging Face), download LLMs directly to your phone, and run inference locally through Google’s LiteRT runtime. It bundles chat, image understanding, voice transcription, prompt engineering tools, and even a tiny garden-building mini-game into one interface.
The interesting bit
The “Agent Skills” system lets you augment models with modular tools—Wikipedia lookup, maps, visual summary cards—loaded dynamically from URLs or community contributions. There’s also a “Thinking Mode” that exposes the model’s step-by-step reasoning chain, though this currently only works with the Gemma 4 family. The README is upfront that this is an experimental beta, which is refreshing.
Key highlights
- Runs inference 100% on-device; no network required after model download
- Supports custom model loading alongside the curated catalog
- Includes benchmarking tools to measure performance on your specific hardware
- “Mobile Actions” and “Tiny Garden” demonstrate function-calling via a 270M parameter FunctionGemma finetune
- Apache 2.0 licensed with APKs available outside Google Play
Caveats
- Thinking Mode is currently limited to Gemma 4 models; broader support is unclear
- Requires Android 12+ or iOS 17+, which excludes older devices that might struggle with on-device LLMs anyway
- Explicitly labeled “experimental Beta release” by the maintainers
Verdict
Worth installing if you’re evaluating on-device AI capabilities or need a private, offline LLM interface. Skip it if you’re looking for a polished consumer product or cloud-scale model access—this is a developer-facing testing ground with app store packaging.