Think out loud.
Nobody else is listening.

Talk through a decision, a hard message or a plan, typed or out loud. The reply is written on your phone, and the words the model was unsure about stay underlined: tap one to see the other words that nearly came next.

A new number calls you Mum and asks for £850. In the same second a warning arrives beside the message, with the reason in one plain line and what to do instead. The check runs on your phone and the link is never opened. If a warning is wrong, Not a scam puts it right with one tap.

“Remind me to call the dentist tomorrow at 9” rings at 9. “Add drinks with Tobe on Saturday at 7pm” goes straight into your phone's calendar, with Undo beside it. When a time could mean two things, I ask before I add anything.

Ask in plain words and I search your messages, journal, chats and documents on your phone. The answer comes back with where it came from, and a tap opens it.

Our own model, built on Qwen3.5 2B, answers, writes, explains and reads photos like any good assistant, and is trained for your phone too: your messages, your calendar, your reminders. Curious 2B runs on your phone like every other model here, and was trained on generated examples and public datasets, never on anyone's chats.