Google has released a free iOS app that runs its Gemma 4 AI model entirely on your iPhone, iPad, or Mac, with no internet connection required after download and no account signup needed. The app, called Google AI Edge Gallery, represents a significant shift in how major tech companies are distributing AI capabilities, moving inference (the computational process of running an AI model) from distant data centers directly onto consumer devices.

What Makes Google AI Edge Gallery Different From Cloud-Based AI?

Google AI Edge Gallery is built on LiteRT-LM, Google's production inference framework designed specifically for running large language models on edge devices, meaning devices at the edge of a network rather than in centralized cloud servers. Once you download a model, your conversations stay entirely on your phone. The app supports multiple sizes of Gemma 4, Google's latest model family released in April 2026, including versions optimized for phones and a multimodal variant that handles text, images, and audio input in a single architecture.

The privacy implications are substantial. While the app does collect some device diagnostics and usage data according to its App Store privacy label, your actual chat prompts and responses never leave your device. This contrasts sharply with cloud-based AI assistants, where every query travels to a remote server.

How to Get Started With On-Device AI on Your iPhone?

• Download the App: Google AI Edge Gallery is available free on the Apple App Store with no subscription fees or in-app purchases required, making it accessible to any iOS user with iOS 17.0 or later.
• Choose Your Model Size: The app offers multiple Gemma 4 variants, including E2B and E4B models designed specifically for phones, a 26B mixture-of-experts variant, a 31B variant, and a multimodal 12B Unified model that processes text, images, and audio.
• Import Custom Models: Beyond Google's built-in models, users can import additional models from Hugging Face, an open-source AI model repository, in LiteRT format, giving power users flexibility to experiment with different architectures.
• Run Offline Conversations: After downloading a model, chat runs completely offline without any internet connection, allowing uninterrupted use even in areas without cellular or Wi-Fi coverage.

Is This App Ready for Everyday Use?

Google itself labels Google AI Edge Gallery as an "experimental Beta" rather than a finished consumer product, a distinction that matters for users deciding whether to rely on it. The app has earned a 4.0 out of 5 star rating from more than 160 App Store reviews as of mid-September 2026, suggesting reasonable stability for early adopters, though the limited review count reflects its recent release.

The app runs on iPhone, iPad, and Mac with Apple Silicon processors, as well as Apple Vision Pro, requiring macOS 14.0 or later for Mac users. This hardware requirement reflects the computational demands of running AI models locally; older devices simply lack the processing power needed for smooth inference.

For users seeking a more polished, actively maintained experience, alternatives exist. Private LLM offers a paid, curated alternative, while PocketPal AI provides a free, open-source option that allows manual import of GGUF format models, giving users different trade-offs between convenience and control.

What Does This Signal About the Future of AI Distribution?

Google's decision to release Google AI Edge Gallery as a free, open-source project signals confidence in on-device inference as a viable path for AI deployment. The underlying project, google-ai-edge/gallery, is licensed under Apache 2.0, meaning developers can freely modify and redistribute the code. The same project also ships on Android and as a direct APK download, indicating Google's commitment to making on-device AI accessible across multiple platforms rather than locking it behind proprietary ecosystems.

The timing matters. As concerns about data privacy, cloud computing costs, and latency continue to mount, major AI labs are investing heavily in making models small and efficient enough to run locally. Google's release of a consumer-facing app, even in beta form, demonstrates that on-device inference has moved beyond research prototypes into practical, user-facing tools. The fact that the app requires no account, no subscription, and no per-message fees removes traditional friction points that have historically limited AI adoption.

For anyone curious about how AI models actually perform when running on consumer hardware, Google AI Edge Gallery offers a rare, free opportunity to test Gemma 4's reasoning and multimodal capabilities without uploading data to Google's servers or paying for API access. Whether you're a developer exploring edge AI, a privacy-conscious user, or simply someone interested in understanding how AI inference works in practice, the app provides a low-risk entry point into the growing world of local AI.