Google has brought AI translation back to its own hands. On August 6, the Google Creative Lab team released an offline translation device called Gemma Translator, which runs its own Gemma4E2B model, with a total of 5.1 billion parameters and 2.3 billion activated parameters. The design goal is clear: specifically designed for resource-constrained edge devices such as smartphones, browsers, and Raspberry Pi.

image.png

The hardware is based on a Raspberry Pi 5 single-board computer. After the user speaks into the microphone, the device converts the speech into the target language, then broadcasts the translation through the speaker, and can also display the original and translated text simultaneously on the screen. The entire processing workflow is completed locally: the translation inference is handled by the LiteRT-LM execution framework prepared by Google for edge device developers, while speech transcription and speech synthesis use Moonshine's AI local processing. This means that even when completely offline, you can communicate face-to-face with people who speak different languages.

Putting a large model into a small development board to run offline translation, the significance of this is not about how large the parameters are, but rather that Google has provided a clear productization path: edge models are no longer just technical demonstrations in laboratories, but can truly be packaged into independent products that are plug-and-play and never rely on the network.