Google AI Edge Gallery: Run Gemma 4 on iPhone With Built-In Tool Calling
Agentic AIGoogle shipped an official iPhone app that runs Gemma 4 models locally — with functional tool calling, image analysis, and audio transcription — and it has received almost no coverage relative to its significance. The app is called Google AI Edge Gallery, and Simon Willison’s hands-on review identifies it as “the first time I’ve seen a local model vendor release an official app for trying out their models” on iPhone.
That framing is the important one. On-device AI inference has been a developer capability for some time. A polished, official, consumer-accessible entry point from a major lab is different.
Context
Running large language models locally on a phone has been technically possible for several years, but the pathway required developer tools, sideloading, or third-party apps with inconsistent quality. The major AI labs have mostly pointed users toward cloud APIs rather than investing in native on-device experiences.
Google’s release of the AI Edge Gallery changes that, at least for the Gemma model family. It positions on-device inference as a first-party product rather than a side project — a meaningful signal about where Google sees the mobile AI stack evolving.
The Detail
The app runs the Gemma 4 E2B and E4B models as well as select Gemma 3 models directly on the device. The E2B (2-billion parameter) model requires a 2.54GB download and performs both quickly and accurately in Willison’s testing. No API calls, no network dependency for inference.
Core capabilities include image analysis (ask questions about a photo), audio transcription for clips up to 30 seconds, and an interactive “skills” demo that demonstrates tool calling across eight built-in widgets:
interactive-mapkitchen-adventurecalculate-hashtext-spinnermood-trackermnemonic-passwordquery-wikipediaqr-code
The tool-calling demonstration is the most significant feature from an agentic perspective. Tool use — where a model invokes functions to retrieve data or perform operations — is foundational to agentic AI workflows. Seeing it run reliably on a 2B parameter model, locally on an iPhone, illustrates how quickly the capability floor is rising. For a technical breakdown of how tool use fits into broader agentic stacks, see Agentic AI Stack Layers: How the Architecture Actually Works.
The app has two notable limitations. First, conversations are ephemeral — no persistent history is saved between sessions, which limits utility for any ongoing use case. Second, Willison encountered a freeze when attempting to add a follow-up prompt in the skills demo. Both are first-version issues, but the conversation log gap in particular would need to be addressed before the app becomes a daily-use tool rather than a demonstration.
The So What
The Google AI Edge Gallery is a proof of capability more than a finished product. But proofs of capability from major labs carry weight. The fact that Google is shipping an official, consumer-facing on-device AI app — with working tool calling — marks a meaningful step toward agentic AI running on personal devices without cloud dependency.
For developers building agentic applications, the implications are direct: smaller models with tool-use capabilities are becoming viable on consumer hardware. The inference cost and latency arguments for cloud-only deployment are narrowing. Architects designing agentic systems in 2026 should be assessing on-device as a serious deployment target, not a future consideration.
Photo by Solen Feyissa on Pexels
Source: Simon Willison
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →