Gemma 4 goes offline in mobile apps: what Google's open-source model actually does without a cloud connection

Video: "Gemma 4 Now Runs Offline In Apps 🤯" by Julian Goldie on YouTube.

What actually changed

Gemma 4 has been available locally via Ollama for a while. What's new is that it now ships inside polished app wrappers on mobile and desktop - not a terminal command, not a configuration file, just an app you open and use. The Google AI Edge Gallery brings Gemma 4N, the mobile-specific variant, to Android as a proper installed application. Gemma Chat handles the Mac side. Neither needs an internet connection once the model is downloaded. In practice, you can run AI on a device with no Wi-Fi and no subscription ticking away in the background.

Google AI Edge Gallery: Gemma 4 on Android

The Google AI Edge Gallery is Google's own Android app for on-device AI. Gemma 4N sits inside it - a variant of Gemma 4 tuned for the constraints of mobile hardware, where memory and battery both matter. It handles basic reasoning and Q&A tasks competently. Context window is shorter than the cloud version, so you're not feeding it a fifty-page document. For the kinds of tasks you actually do on a phone - drafting a reply, summarising a short brief, writing a quick product description - it performs better than you'd expect from something running on a handset with no connection.

Worth knowing: Gemma 4N is not the same model as the desktop version. It trades raw capability for the ability to run on constrained hardware. Most businesses will find the output useful for short-form work and limited for anything that requires extended reasoning.

Gemma Chat on Mac: building apps from prompts

The desktop version is the more interesting development. Gemma Chat runs entirely on your Mac and accepts plain-English prompts to build small web tools - landing pages, calculators, games, forms, quizzes. The workflow Julian Goldie demonstrates is straightforward: you type what you want, a live preview appears, and you keep refining the brief until the output matches. Nothing is being sent to a cloud server. The finished output is an HTML file you can drop onto any hosting platform.

That said, be realistic about what "build an app" means here. These are simple, single-purpose tools, not full web applications. A calculator, a lead capture form, a countdown page - yes. A multi-page site with a backend and user accounts - no. For quick internal tools and throwaway marketing pages, the capability is real. For anything more complex, you still need a developer or a frontier model with proper coding support.

Ollama 0.31 and the Apple Silicon speed jump

If you're already running Ollama on an M-series Mac, version 0.31 is worth updating to immediately. It brings a 90% speed improvement for Gemma 4 on Apple Silicon - the practical difference between generation that feels sluggish and one that feels usable for iterative work. Pull the updated model tag in Ollama, restart, and the speed is there. Nothing to configure.

The improvement is specific to Apple Silicon's architecture. Intel Mac users and Windows users running Ollama will see smaller gains. Linux on ARM hardware should benefit similarly, but Julian Goldie's walkthrough focused on the Mac side.

What's genuinely worth your attention

Gemma 4 offline is not competing with GPT-5.6 or Claude Fable 5. It is a capable, free, open-source model with a privacy advantage: nothing leaves the device. The sweet spot is tasks where you'd otherwise spend a few pounds a month on cloud tokens for routine work - drafting, summarising, building simple tools. On sensitive internal data you don't want uploading anywhere, running a local model is the right choice regardless of cost.

To be fair: most UK businesses are not yet running any local AI model, let alone Gemma. The gap between "I've heard of Ollama" and "I have a working setup that saves me an hour a week on low-stakes tasks" is where the value is sitting. The new app wrappers make that gap smaller to close.

Where this connects to NordSys

If you want to build internal tools or automate routine tasks while keeping AI costs low and your data on-site, combining a local model like Gemma 4 with cloud models for heavier work is a practical setup. We configure local AI stacks - Ollama, Gemma, Hermes Agent - alongside the frontier models, so you use the right tool for each task rather than paying for Opus-class compute on work that a local model handles fine.

See our Programming service →