On-device model goes live; steadier, snappier computer use

On-device model

• Small features — titles, screen perception, typing completion, and memory upkeep — now run on the device's own model by default: nothing leaves the device, no extra cost, and faster responses (needs the device updated to the latest version).
• Settings → AI feature usage gains a switch and a marker: see at a glance which features run on the on-device model, switch any of them back to the cloud or to a specific model, and see roughly how much the on-device model has saved you.

Jev fast decisions: faster, steadier computer & web use

• New Jev fast decision loop: when operating your computer or the web, the Jev decision model looks at the screen, picks the next step, and runs a multi-step goal on its own (look → pick one step → act → repeat) — much quicker than stepping by hand, and lighter on usage.
• For the few apps that ignore background control, the AI briefly comes to the foreground, uses real keyboard/mouse, and restores your focus when done; regular apps still run in the background without taking over your foreground.
• The click ripple is now a smoother native animation.
• Fixed cases where pressing Enter in some chat input boxes only added a line break instead of sending.
• Typing is now visible — you can watch the AI type character by character.

Completion & interaction

• Typing completion triggers more responsively and pre-warms the current field; it recovers on its own when it occasionally fails to appear.
• After a tool finishes, the current note stays put instead of being instantly replaced by "Thinking…".

Models & performance

• Added OpenRouter as a model source (with its icon and selectable models).
• Less background repainting and fewer needless reconnects — long sessions use fewer resources.

Privacy

• Fewer repeated screen-recording permission prompts.