Notes on private, on-device assistants: how they work, how to check one yourself, and what changes when the language model runs on your phone instead of on a server.
"Private" is doing three different jobs on app pages right now: a promise, a setting, or an architecture. Seven checks, each under a minute, that tell them apart on your own phone.
Which phones can do it, what to download, what to expect from speed and quality, and the one test that proves nothing is leaving.
What the model is, why a four-billion-parameter model is the size that fits a phone, how fast it runs with MLX, and what it can and cannot do offline.
Not a comparison of servers. A comparison of what happens to your words, how fast the answer arrives, what it costs, and how much the assistant can do.
If the reason is privacy, the question is not which chatbot to switch to. It is where the model runs, and what you gain and give up when it runs on the phone.
Orvena speaks 31 languages in voices generated on the iPhone, and its interface is in 30. Why local matters more for voice than for text, and how it works.