On May 31, 2025, TechCrunch reported that Google had quietly released an app called AI Edge Gallery, which lets users download and run AI models directly on their phones with no internet connection required. There was no launch event and no press release; the code simply appeared on GitHub, a notably low-key move for a company that markets its cloud AI services aggressively.
AI Edge Gallery is available now as an experimental Alpha release for Android, with an iOS version promised soon. Its core capability is pulling open models, including Google’s Gemma family, from Hugging Face onto a device, after which all inference runs locally. For developers who care about privacy, work in poor-connectivity environments, or simply want to avoid cloud inference costs, the app sends a clear signal: Google is treating the phone as an inference device as a real product direction.
A quiet experimental release
According to TechCrunch’s May 31 report, the app went live earlier that week as an experimental Alpha under the Apache 2.0 license, with source code published in Google’s google-ai-edge repository on GitHub. The repository describes itself as a gallery that showcases on-device ML and GenAI use cases and lets people try and use models locally. Compared with the marketing push behind Gemini, this release got essentially no promotion, which suggests Google itself framed it as an experiment: let developers kick the tires first, then decide which parts deserve to become products.
Three surfaces: AI Chat, Prompt Lab, and Ask Image
The app ships with three main interfaces. AI Chat is a general-purpose conversation surface for multi-turn chat with a loaded language model. Prompt Lab is built for single-turn tasks: it comes with prompt templates and adjustable settings, which makes it a handy scratchpad for testing how summarization or rewriting instructions behave on a small model. Ask Image supports multimodal queries, letting users ask questions about a picture. Supported models include Google’s Gemma 3n, designed for on-device use, plus other compatible models pulled from Hugging Face, covering question answering, code writing and editing, and image understanding. For developers, it amounts to an officially maintained playground for comparing edge models without assembling an inference stack themselves.
Why offline inference matters
Once a model is downloaded, inference happens entirely offline, which buys three concrete benefits. Privacy: conversations and images never leave the device, which matters for internal tools and sensitive data. Availability: the app keeps working on flights or in areas with weak signal, with no dependency on an API service’s uptime. Cost: every inference is free, which makes the app an affordable testing ground for prototypes that would otherwise rack up API bills. It also lowers the barrier to experiencing on-device model behavior firsthand; until now, comparing open models on a phone meant building and wiring an inference framework by hand.
Hardware limits and uneven performance
Google’s own documentation warns that speed and output quality vary significantly with device hardware and model size. A flagship phone can run small models comfortably, while mid-range devices may face slow generation or fail to load larger models at all due to memory limits. That is the standing trade-off of edge AI: the gap between model capability and device compute cannot be wished away, so model selection has to weigh parameter size against the device distribution of your target users. For product decision-makers, the app’s real value is as a low-cost evaluation tool for deciding whether a given use case belongs on-device, in the cloud, or somewhere in between.
Google’s broader edge AI play
Context makes the release more interesting. In 2025, Google was pushing Gemini hard in the cloud for subscriptions and API revenue while also investing in the open Gemma model family and the AI Edge toolchain; AI Edge Gallery extends that line straight onto consumer phones. The Apache 2.0 license and fully open source code indicate Google wants to win through an ecosystem rather than a single app: developers can inspect, modify, and extend it, and the model catalog is not locked to Google’s own. Looking back from 2026, this quiet release marks one of the turning points where offline mobile AI moved from demo to practical tool, and it is a reasonable starting point for tracing the on-device model competition that followed across mobile platforms.
Sources
- TechCrunch: Google quietly released an app that lets you download and run AI models locally
- GitHub: google-ai-edge/gallery
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.