What can and cannot work without a network
The model runs on a server, so answers need a connection. Offline support in an AI app means the interface loads, past conversations are readable, and a queued message sends when you reconnect. Useful, and not what the word suggests.
Offline is claimed often and means something narrow here. The distinction matters because the useful part of the product is precisely the part that cannot go offline.
The model is the reason
Generating a reply needs the model, and a capable model is far too large for a phone. Some small models run on devices and they are noticeably weaker. This is a physical constraint rather than a product decision.
What offline support does provide
The interface opens instead of showing an error, past conversations stored on the device are readable, and a message typed offline can be queued. These are real conveniences and they belong to the shell rather than to the assistant.
How to tell what a product means
Turn off the network and ask something. A queued message with a clear pending state is honest offline support; a spinner that never resolves is an app that did not consider it. This is a 30-second test.
On-device models, briefly
Running a model locally is real and improving, and today it means a much smaller model, a large download and a warm phone. For short factual work it is usable; for the work people bring to an assistant it is not there yet.
Questions about the app
Would an offline model be more private?
Substantially, because nothing leaves the device. It is also weaker, and the trade is the honest way to frame it.
Why do apps advertise offline at all?
Because the shell working offline is genuinely better than an error page. The wording oversells a real improvement.
Does a web app handle this worse?
No. The same mechanism that caches an installed app caches a web app, and both depend on the same server for answers.
Turn off the network and see exactly what still opens.
Open the chat first