It’s gotta be just barely holding on in some legacy environment right now. I’m surprised they don’t just start over using new tech. Maybe that’s the end goal with on-device inference to sunset cloud Siri.
Otherwise why don’t they let older devices use their server-side private LLM setup?