Google built a portable AI translator that works offline, and you can build one too

Google’s Gemma models are quite capable, and if you’ve been wondering exactly how good these models are, there’s something you need to see. A team of Google engineers has created a fully offline translator that uses the Gemma 4 E4B model running on a Raspberry Pi 5.

The team used Google Antigravity for development and the Gemma 4 E4B model for the AI chops needed to translate different languages. They also designed a 3D-printed housing to fit all the electronics, complete with a display to show the translated text, a mic, a speaker, a push-to-talk button, and a knob to change languages.

The team used LiteRT-LM to run the Gemma 4 E4B model and Moonshine text-to-speech to output the translated speech via the speaker. Everything runs on a Raspberry Pi 5, and there’s a tiny touchscreen with a custom-built interface to display translations in real time.

The best part about the entire project is that it runs entirely offline after the initial setup. So, it can be used anywhere without cellular or Wi-Fi connectivity. It also shows how capable the Gemma 4 E4B model is at handling translations without connecting to a server for extra processing power. Most translation apps, such as Google Translate, require an internet connection to translate speech.

The translator is an excellent showcase of both Google Antigravity and the Gemma 4 models. All the code and 3D-printing files for the enclosure are available on GitHub for anyone to download, compile, and create their own offline translation device.