Google’s LiteRT toolkit lets a Raspberry Pi 5 run compact Gemma language and vision models locally, without sending inference to a cloud service.