Product
Six models that never phone home
Offline AI shipped this spring. Here's how we picked the models, and what it means to run them on a phone with no signal.
An assistant that stops working when the signal does is not much of an assistant. So the offline AI in Nobars is exactly that: open models, downloaded once, running on your device, answering in a tunnel at the bottom of a valley on a plane with the Wi-Fi off.
How we chose
We wanted a range, not a winner. The quickest model answers in under a second on a four-year-old phone and is happy translating a menu or drafting a message. The deepest one takes its time and is worth it when you are trying to work through something. All six are open weights, so we can tell you exactly what you are running.
- Quick: sub-second answers, about 400MB, fine on older phones.
- Everyday: the default. Balanced, about 1.2GB, handles most questions and rewrites.
- Deep thinker: slower, more careful, for when you have a minute and a hard question.
- Translate: 40 languages, tuned for short conversational text.
- Summarise: built for long notes and transcripts, not chat.
- Code: for the people who take laptops camping.
The part people don't expect
The most used feature has not been questions. It is rewriting. People draft a message they cannot quite phrase, ask for a shorter version, and send it over the mesh to someone forty metres away. The whole loop — the thinking and the sending — happens without a single packet leaving the field.
Battery is the honest trade. A long conversation with the deep model is the most expensive thing Nobars can do, and we show you that cost in the app rather than hiding it. In practice most people keep Everyday selected and never think about it again.
