Google AI Edge Gallery Review: Gemma 4 On-Device Chat for iPhone

Google AI Edge Gallery is Google's own showcase app for running open-source generative AI models directly on a phone, tablet, or Mac, with no server round-trip for inference. It is built on LiteRT-LM, Google's production inference framework for large language models on edge devices, which itself succeeded the older MediaPipe LLM Inference API used by earlier versions of the project.

The iOS listing reviewed here is Google AI Edge Gallery on the Apple App Store (App Store ID 6749645337), developed and published by Google LLC, currently at version 1.0.13 and rated 4.1 out of 5 stars from 165 ratings. It is a free download with no subscription tier.

This same app is one distribution channel of a larger open-source project, google-ai-edge/gallery, which Google describes as "the premier destination for running the world's most powerful open-source Large Language Models (LLMs) on your mobile device." That project also ships an Android build (via Google Play and direct APK download) and a macOS DMG. Google's own project documentation labels the whole effort an "experimental Beta release" rather than a finished consumer product — a distinction this review treats as material, not a formality, given that Google AI Edge Gallery is compared throughout this cluster against apps built specifically as polished consumer products.

Google AI Edge Gallery is free to download, with no purchase step. This review's CTA covers the iOS listing specifically; the Android and source-code channels are separate downloads from the same underlying open-source project, not the same install.

This review is a companion to Google AI Edge Gallery's entry in the Local LLM Software Directory, which lists it alongside every other mobile local-AI app covered on this site.

Google AI Edge Gallery centers on Google's Gemma model family, currently the Gemma 4 generation released April 2, 2026. Gemma 4 ships in multiple sizes — E2B and E4B (designed for phones), a 26B mixture-of-experts variant, a 31B variant, and a multimodal Gemma 4 12B Unified model that handles text, image, and (per Google's own documentation) audio input in a single architecture. The app also supports Gemma 3n, an earlier Gemma release, which added audio understanding to the on-device stack before Gemma 4 shipped.

Beyond the bundled Gemma models, the app lets users import their own models from Hugging Face in LiteRT format, using a model-card URL inside the app's model-management screen. Google's own project documentation confirms this Hugging Face integration exists for model discovery and download, but does not spell out details like whether a Hugging Face access token is required for gated model repositories — this review treats that as unconfirmed rather than assuming a specific token flow, and readers should check the current in-app instructions before attempting to import a gated model.

Feature-wise, the app packages several distinct modes around these models: AI Chat with Thinking Mode (which surfaces the model's intermediate reasoning steps, starting with the Gemma 4 family), Ask Image (multimodal analysis of a photo from the camera or gallery), Audio Scribe (on-device speech transcription and translation), Agent Skills (external tools such as Wikipedia lookups, interactive maps and visual summary cards, layered on top of the base model), Instant Media Search and Video Moment Finder (natural-language search of photos and videos, powered by EmbeddingGemma 2), Prompt Lab (single-turn prompt tests with temperature and top-k controls), Mobile Actions and Tiny Garden (offline device controls and an experimental mini-game, both built on a FunctionGemma 270m finetune), and a model management and benchmarking screen for comparing download size and on-device performance across models before committing to one. The listing and README describe the image features as analyzing or searching photos you already have; none is described as creating new images.

Inference happens on-device once a model is downloaded — chat prompts and model output are not sent to a Google server to generate a response. That said, this is not the same claim as "the app collects no data," and this review treats those as two separate statements rather than conflating them.

Apple's privacy nutrition label for the Google AI Edge Gallery App Store listing states that data linked to your identity includes device ID, performance data, diagnostics, and unspecified other data types, and that data not linked to your identity includes coarse location, product interaction data, and crash reports. This is Apple's standard, developer-declared privacy label format, not an independent PromptQuorum audit of the app's network traffic.

The practical distinction for a privacy-conscious reader: the content of your conversations and any images or audio you feed the model stay on the device during inference, but Google does collect telemetry-style data (device identifiers, performance/diagnostic data, coarse location, crash reports) about how the app itself is used. Readers who need a stricter "collects nothing at all" posture should compare this against Private LLM's App Store privacy label, which states no data is collected from that app.

Google AI Edge Gallery began as a GitHub-only open-source project before reaching app stores. The timeline below is sourced from Google's own blog posts and the project's GitHub repository.

Google AI Edge Gallery earns a place in this cluster as the fastest, free way to run Google's own Gemma 4 models entirely on an iPhone, backed by Google's production LiteRT-LM runtime rather than a third-party wrapper around an open-source model. Its Hugging Face import support and multimodal features (Ask Image, Audio Scribe, Thinking Mode) go beyond a bare-bones chat client. The trade-offs are equally real: Google itself labels the project an "experimental Beta," the iOS App Store privacy label shows more data collection than the strictest privacy-focused competitors, and the app has a shorter track record as a maintained consumer product than apps like Private LLM. Readers who specifically want the newest Gemma models at no cost, or who are evaluating LiteRT-LM as developers, should install it. Readers who want the most polished, longest-running, or most privacy-restrictive mobile local-AI app should compare it first against Private LLM or PocketPal AI.