Google's offline AI app

Google has brought its AI Edge Gallery app to the iPhone, giving developers and testers a way to run some generative artificial intelligence features directly on the device instead of in the cloud. (developers.googleblog.com) Google’s AI Edge team said in a March 2026 blog post that the iOS version joins Android in letting people try “high-performance, on-device AI use cases” with Gemma and other open-weight models. The app is distributed through Google’s AI Edge project on GitHub, where Google describes it as a showcase for running models locally on mobile hardware. (developers.googleblog.com) (github.com) Running a model “on-device” means the phone does the processing itself, like editing a photo on the handset instead of sending it to a remote server. Google says AI Edge Gallery is built for that setup, with offline use, local execution and model testing on the user’s own hardware. (github.com) (ai.google.dev) The timing is tied to Gemma 4, the open-model family Google DeepMind launched on April 2, 2026. Google said the smaller Gemma 4 E2B and E4B models were designed for edge devices, with multimodal input support, low-latency processing and deployment across phones, laptops and other local hardware. (blog.google) (ai.google.dev) Multimodal means one model can handle more than one kind of input, such as text and images, instead of only typed prompts. Google’s Gemma 4 documentation says the models can take text and image input and generate text output, with audio support on the smaller edge-focused versions. (ai.google.dev 1) (ai.google.dev 2) Google has been using AI Edge Gallery as a test bed for features that normally depend on bigger cloud systems. In April 2026, the company said the app now includes “Agent Skills,” which it described as one of the first applications to run multi-step autonomous workflows entirely on-device. (developers.googleblog.com) That work follows an earlier iPhone update centered on on-device function calling, a technique that lets a model trigger tools or actions inside an app instead of only returning text. Google said that release brought AI Edge Gallery to iOS so developers could explore the same local model workflows available on Android. (developers.googleblog.com) Google is also pushing the same local-model strategy across Android. The company said Gemma 4 is available in Android’s AI Core developer preview, where it advertises faster inference, lower battery use and support for more than 140 languages on supported devices. (developer.android.com) (developers.googleblog.com) The app does not replace Google’s cloud-based Gemini products, which still handle many of the company’s largest consumer AI features. It does show Google trying to make a second lane for artificial intelligence on phones: smaller open models, running locally, with no network connection required for at least some tasks. (github.com) (blog.google)