Google's AI Edge Gallery app brings offline AI to your phone — but hardware support lags

At a glance:

• Google AI Edge Gallery, an experimental app updated to support Gemma 4, lets users run open-source AI models entirely on-device — no internet required.
• Three standout offline features: an AI chatbot with multimodal input (text, voice, images), an audio scribe that transcribes and translates speech in real time, and an "ask image" tool for querying photos.
• The app runs well on iPhone (leveraging GPU) but on Android — even the Pixel 10 Pro — it falls back to CPU processing because Tensor GPU/NPU support is locked behind a beta program, making responses 10× slower.

What is Google AI Edge Gallery and why it matters

Google AI Edge Gallery is not a brand-new product. Launched roughly a year ago as an experimental app, it has quietly sat in the Play Store and App Store gathering dust for many users. What changed the equation is an update that adds support for Gemma 4, Google's newest and most capable open-source AI model family. That single addition is what prompted reviewers and testers to give the app a serious look — and what makes the story worth telling now.

The core premise is straightforward: you download open-source AI models directly onto your phone and then run them locally for a range of predefined tasks. No cloud round-trip, no API key, no data leaving the device. The app is cross-platform, available on both Android and iOS, and the feature set is consistent across operating systems even if the underlying hardware utilization differs dramatically, as we will see.

For a long time, on-device AI has felt like a marketing checkbox more than a practical tool. Most AI features on phones still lean on the cloud or a hybrid architecture, and the processing power of a smartphone chip has not seemed up to the task of running anything beyond toy-scale models. Google AI Edge Gallery, powered by Gemma 4, is one of the first examples that genuinely challenges that assumption in daily use.

Three offline AI features that actually help

The app ships with several predefined use cases, and three of them proved surprisingly useful during real-world travel:

• AI Chat — a general-purpose chatbot similar in spirit to Gemini or ChatGPT. It is multimodal, so you can feed it text, voice, or images and it will treat all of that as context. Responses are slower than a cloud-backed chatbot, but the fact that everything runs locally with zero internet is the differentiator.
• Audio scribe / offline translation — a dedicated tool that transcribes speech and translates it on the fly. On phones that properly tap into the hardware, translation speed is close to what you would get from a dedicated translation app, and it works reliably with no connectivity.
• Ask image — you attach a photo and ask anything about it. The reviewer found this especially handy for translating restaurant menus or deciphering street signs in foreign languages while abroad.

Each of these features saves mobile data because nothing is uploaded to a server. That alone is a meaningful benefit for travelers or anyone on a limited data plan.

A real-world test: 32,000 feet with no internet

The reviewer put the app through its most dramatic stress test on a flight to Thailand. With no internet available at cruising altitude, they opened AI Chat and asked practical questions: useful Thai phrases, movie titles shown on the in-flight entertainment system, and approximate IMDb ratings for recommendations. The model correctly stated that it cannot access the internet and relies only on its training data, but when questions were framed around static knowledge, the answers were accurate and helpful.

The offline translation feature proved equally valuable. The reviewer travels with an eSIM but frequently encounters areas with poor connectivity. Being able to point the phone at a speaker and get a near-real-time translation — without any data connection — made the app feel like a genuine productivity tool rather than a gimmick.

The "ask image" feature completed the trio of surprises. While the reviewer did not expect to use it much, snapping a photo of a menu or sign and asking the model to explain or translate it saved both time and data. Everything stayed on the device.

Hardware reality: iPhone vs Pixel performance gap

Here is where the story takes a sharp turn. The app's behavior differs sharply depending on the platform and chipset.

On iPhone, the app leverages the GPU for AI processing, which is significantly faster for neural-network workloads. On Android, the experience is inconsistent. Phones with top-tier Snapdragon chips — specifically the Snapdragon 8 Elite Gen 5 — can tap into the GPU and run the app efficiently. But on Google's own Pixel 10 Pro, the app does not appear to fully utilize the Tensor GPU or the chip's NPU.

The reviewer explains that the AICore-based experience, which would tap into the NPU, is currently limited to beta testers. For everyone else, the app falls back to CPU processing. That fallback is noticeable: on the same audio input, the iPhone Air responded in under one second, while the Pixel 10 Pro took more than ten seconds. A tenfold gap in latency undermines the entire offline-AI value proposition.

This is not a small complaint. Google has been one of the loudest voices pushing on-device AI, and the Pixel line is its flagship Android showcase. Seeing the Pixel 10 Pro — a phone with a dedicated Tensor chip and an NPU — run the app on CPU while an iPhone uses its GPU makes the inconsistency hard to ignore.

What the app still needs to fix

Beyond the hardware utilization issue, there are two other pain points the reviewer highlighted:

• Chats are not saved. Unlike Gemini or ChatGPT, which store each conversation as a thread you can return to, AI Edge Gallery does not persist chat history. The reviewer acknowledges that offline models have context-length limits due to hardware constraints, but argues there should at least be an option to continue a conversation up to that limit.
• Android hardware support is incomplete. The app's reliance on CPU on flagship Android phones — even ones with Tensor or Snapdragon 8 Elite Gen 5 — means the experience lags behind what is technically possible. Until AICore support expands beyond beta, the Pixel 10 Pro and similar devices will feel sluggish compared to iOS counterparts.

The reviewer notes they may be in the minority talking up an app with limited capabilities and a niche user base, but given how aggressively Google has marketed on-device AI, the gap between promise and execution is frustrating.

Who should try it and what to watch next

If you travel frequently, work in areas with unreliable connectivity, or simply want a private AI assistant that never sends your data to the cloud, Google AI Edge Gallery is worth a download. The app runs on Android and iOS, has been tested on the iPhone Air, Google Pixel 10 Pro, and Oppo Find X9 Ultra, and the three core features — AI Chat, offline translation, and ask image — cover most offline-AI use cases.

What to watch: Google's rollout of AICore support beyond beta testers could close the performance gap on Pixel and other Android flagships. If and when the app fully utilizes the Tensor NPU on Pixel devices, the ten-second-vs-one-second disparity should narrow dramatically. Until then, iPhone users get the noticeably snappier experience.

The bigger takeaway is that on-device AI has finally crossed the line from gimmick to genuinely useful — but only if the hardware support matches the software ambition.

Key takeaways

• Google AI Edge Gallery supports Gemma 4 and runs entirely offline on-device.
• Three core features — AI Chat, audio scribe/translation, and ask image — proved practical during real-world travel with no internet.
• Performance is strong on iPhone (GPU) but sluggish on Pixel 10 Pro (CPU fallback) because AICore/NPU access is beta-only.
• Chat history is not saved, and Android hardware utilization remains incomplete.

Tags

• google-ai-edge-gallery
• gemma-4
• on-device-ai
• pixel-10-pro
• offline-ai
• ai-edge-gallery