• Open source
  • Story
  • Feb 26, 2026

Google's Edge Gallery runs AI on your phone, in airplane mode

My take: An AI that still answers in airplane mode is one I can trust with a photo of a contract. I haven't run it yet: this is my reminder.

What to know

  • Google AI Edge Gallery is an open-source app (Apache 2.0, 25,100 stars on GitHub on October 26, 2026) that runs language models on the phone itself. It needs Android 12 or iOS 17 and up, and Google still calls it an experimental beta.
  • Once a model is downloaded, no connection is needed to chat, ask about a photo, transcribe or translate a voice recording, or test prompts. The models, such as Google's Gemma family, come from Hugging Face, and you can load your own.
  • On February 26, 2026, iOS joined Android, along with FunctionGemma, a 270-million-parameter model that turns a sentence like "Turn on the flashlight" into a phone action, offline. Google measured it at 142 tokens (word pieces) per second on a Pixel 7 Pro's processor.

Spotted on Korben's blog, in French.

What changes

For what you would rather not send to a server, such as a contract or a client's photo, an AI that stays on the phone is now one install away. The models that fit on a phone are smaller than the ones in the cloud: try yours on a real task before you count on it.

Source: Alice Zheng, Milen Ferev and Wai Hon Law, Google, “On-Device Function Calling in Google AI Edge Gallery”, February 26, 2026

Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.

Up next