Smaller language and vision models can now run on consumer laptops and phones. That makes instant assistance possible in places where connectivity, cost or privacy would otherwise be a limitation.
Local processing does not solve every privacy concern, but it gives developers another option. Sensitive notes, photos and recordings can be analysed without automatically leaving the device.
The trade-off is capability. Cloud models remain stronger for many complex tasks, so the most useful applications will choose carefully between local speed and remote scale.