Rumi: a language model that answers on your own phone
Rumi downloads a language model onto your phone and runs it there. When you type a question, the answer is computed on the device rather than sent away to be answered.
Rumi used to be described as an app that makes no network requests. That was not true, and the claim has been withdrawn. It signs you in, it downloads models, and the free tier shows ads. What it does not do is send your conversation anywhere, and the difference between those two statements is the whole point of this page.
Where it stands
In development and internal testing. Not yet available in the app stores; when it ships, the store links will appear on this page.
Answers computed on the device
The model file lives in the app's own storage and runs there. The text of your conversation is not sent to a server to be answered.
One door for the network
All network access goes through a single chokepoint that allows only a plain file download from a fixed list of hosts. Adding any other kind of request fails the build.
Models from Hugging Face
Weights are downloaded from six named Hugging Face hosts, with no account identifier and no cookie attached to the request.
Conversations stay on the phone
Chats and notes are files in the app's private directory. There is no Rumi server and no account record beyond the sign-in.
What you type is answered by a model running on your phone and is not transmitted for an answer.
Rumi names its own exceptions rather than hiding them: signing in reaches Google, model downloads reach Hugging Face, and ads reach Google. Its source code says the stronger marketing claim is unsupportable.
Want in before the stores?
We invite a small group of early testers per app. If you would use Rumi for real, tell us below and we will set you up when the next test round opens.
Rumi is one of the apps in the Deplyra portfolio. It is designed, built and run by the same team you can hire: AI product engineering is what this app is proof of.