Google Creative Lab ships offline Gemma translator for Pi 5

A repository from Google Creative Lab that landed on GitHub is worth a look, not because it is a flagship model release but because of what it demonstrates about where on-device inference has actually reached. The Gemma Translator project packages a fully offline voice translation setup that runs on a Raspberry Pi 5 with 8GB of RAM, using the gemma4-e2b model via LiteRT-LM for the translation step and Moonshine for both speech-to-text and text-to-speech.

The stack itself is the interesting part. A React and Vite front end tuned for a 480x320 display, a Python http.server backend, and a language model running entirely locally after setup. No cloud call in the loop for the translation. The team credits Alan Yam, Shashwath Santosh, and Dan Motzenbecker of Google Creative Lab, and the code ships under Apache 2.0.

The honest caveat is that the README is thin on the things you would actually want to know before betting on it. There is no specific list of supported language pairs, no published quality or latency numbers, and the repo carries the standard 'not an officially supported Google product' disclaimer, so treat it as a reference build rather than a supported platform. What the documentation does not give you is real-world benchmark quality against hosted translation APIs on longer, noisier or domain-heavy utterances.

For teams thinking about privacy-sensitive or air-gapped translation, medical intake, legal work, on-set film, field research, a working blueprint that runs on commodity Pi-class hardware is more useful than another leaderboard chart. That is the part worth watching.