PRODUCTION SCENARIO
A field-service company needs a small generative model on technicians' rugged tablets, which work with no network connection. Its ML team also wants to tune the model with its own PyTorch pipeline.
Which option fits both constraints?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Gemma, downloaded to the tablets
Gemma is a family of lightweight open models built from the same research as Gemini, and Google documents them as available to run in your applications and on your hardware or mobile devices. Open weights are what let the ML team tune the model in its own PyTorch pipeline. Running Gemma as a managed endpoint is the near miss worth knowing: right model family, wrong deployment, because a hosted endpoint still needs the network the tablets do not have. Flash-Lite is small and cheap but is a managed model reached over the network, so 'Lite' is not the same as on-device. Caching a Gemini model's responses serves answers already seen; a new question on a disconnected tablet has nowhere to go.
Official Source: https://ai.google.dev/gemma/docs ↗