Running AI models directly on the iPhone or Mac’s neural engine means personal data like photos, messages and voice recordings often never leave the device for a feature to work.
This approach trades some raw model size and capability for privacy and speed, since on-device models are constrained by the hardware’s memory and power budget compared to massive cloud-based models.
For tasks like photo search, dictation and predictive text, this trade-off is barely noticeable to users while meaningfully reducing what data is exposed to outside servers.