Google's new app lets you run Hugging Face models offline on your phone

Running AI models directly on your phone usually meant either wrestling with complex developer runtimes or giving up and sending all your sensitive data to cloud servers. Google wants to cut out the middleman entirely.

Google quietly rolled out an open-source app called Google AI Edge Gallery. The app allows users to find, download, and run a variety of open-source models directly on mobile devices without an internet connection.

Built on Google's LiteRT runtime and licensed under Apache 2.0, the app runs completely on-device. That means zero cloud roundtrips for your private prompts, documents, or photos. While initially released as an alpha on GitHub, it is now available across Android, iOS, and macOS.

A local sandbox for open models

The app bypasses bare-bones terminal setups and gives users a polished interface packed with practical everyday tasks.

For voice conversations, AI Chat integrates on-device speech recognition and text-to-speech, letting you talk to local models out loud. The Ask Image tool taps into multimodal models like Gemma 3n so users can point their camera at objects or feed in photos to solve visual questions.

Thanks to newly added support for EmbeddingGemma 2, the app also packs on-device multimodal semantic search. Features like Instant Media Search and Video Moment Finder let you search your photo library or scrub through local video timelines using natural language descriptions.

For tinkerers, a dedicated Prompt Lab allows fine-tuning model parameters like temperature and top-k for single-turn text tasks. There is even a mini-game called Tiny Garden and a Mobile Actions tool, both driven offline by a finetuned FunctionGemma 270m model.

Performance still hinges on your phone hardware. Google notes that larger models tackling multimodal tasks will predictably take longer, and you will need at least Android 12 or iOS 17 to run the app.

Why it matters

For developers and privacy-minded users, this app turns mobile hardware into an instant local AI testbed. You no longer need to write boilerplate code or set up custom runtimes just to benchmark how an open model behaves on a smartphone.

For the industry, it signals Google's aggressive push to make on-device inference practical and mainstream. Serving frontier cloud models is ruinously expensive, so pushing everyday lightweight tasks directly onto edge hardware helps Google relieve cloud infrastructure strain while expanding the reach of its Gemma family.

We think releasing this under an Apache 2.0 license across both Android and iOS is a remarkably smart distribution play. Instead of locking local AI features behind Pixel exclusivity, Google is using Hugging Face integration and open access to establish LiteRT as the default edge execution layer.

The real challenge, however, will be thermal limits and battery drain. Running local weights on mobile processors quickly hits physical boundaries, and if everyday devices throttle within minutes of heavy use, local AI will struggle to break out of niche developer circles.