Fixes the first thing you see. The app runs its AI model on your iPhone, so it has to download that model once before it can answer anything. It used to start that download silently, behind your first question — so a normal one-to-three-minute wait looked exactly like the app had frozen. Now it tells you. The opening screen asks you to download the model and says how big it is, the status reads "No model yet" until it's there, and the download shows a live percentage and elapsed time. If your connection drops it says so, and nothing is lost — asking again picks up where it left off. You'll also see the data-sharing screen once more. It now names the model download alongside everything else, so the list of what leaves your device is complete.

Model downloads now show real progress. The progress bar used to sit at a few percent for minutes on a perfectly healthy download, then jump to 100% — it now moves with the bytes actually received. Loading a model into memory is now its own labelled step, so picking a model you've already downloaded no longer shows a full, motionless bar while it loads. The recommended on-device model is now Qwen3.5 0.8B.