There is no use case for local models.

See Gemini Nano. It is available in custom apps, but the results are so bad; factual errors and hallucinations make it useless. I can see why Google did not roll it out to users.

Even if it was significantly better, inference is still slow. Adding a few milliseconds of network latency for contacting a server and getting a vastly superior result is going to be preferable in nearly all scenarios.

Arguments can be made for privacy or lack of connectivity, but it probably does not matter to most people.