Apple foundation models

What it is

Apple Intelligence models are the foundation models Apple has built to power Apple Intelligence, the set of AI features woven through its devices, and they are notable for a strategy that differs from most of the industry: much of the work happens on the device itself. Apple has been a quiet presence in the race to build large models, saying less and demonstrating fewer headline systems than its rivals, yet it reaches an installed base measured in billions of devices. That combination, restraint in public and enormous distribution in private, is what makes the Apple Intelligence models worth understanding.

The on-device choice

The defining choice is on-device processing. Rather than send every request to a data centre, Apple runs a capable model directly on the phone, tablet or computer in the user’s hand. The appeal of on-device ai is concrete. Work done locally does not require a network connection, so it functions offline and returns results without the delay of a round trip to a server. It is cheaper to run at scale, because the user’s own hardware does the work in place of a rented fleet of servers. Above all it is private, since data processed on the device need never leave it, which fits a company that has long treated privacy as part of its pitch.

Where the cloud comes in

On-device models face a hard limit, though, because a phone cannot hold or run the very largest systems. Apple’s answer to the tasks that exceed what local hardware can manage is Private Cloud Compute, a server tier designed to extend the same privacy guarantees to the cloud. The promise is that when a request is too demanding for the device, it can be handled on Apple’s servers under conditions the company says are built so that even it cannot see the user’s data, and that outside experts can inspect. This two-part structure, a model on the device for most work and Private Cloud Compute for the rest, is the architecture underpinning the Apple Intelligence models.

The strategy suits Apple’s position better than it would suit most rivals. A company that sells the hardware can design its silicon and its models together, so that the chips in its devices are built to run the models efficiently and the models are built to fit the chips. That vertical control lets Apple pursue on-device AI where a company without its own hardware would have little choice but to route everything through the cloud. It also means the Apple Intelligence models reach users not as a separate app to seek out but as features already present across a vast installed base, a distribution advantage no amount of marketing can buy.

On-device against the cloud

The wider significance of the Apple Intelligence models is what they say about where AI computation should sit. Most of the industry has bet on the cloud, on ever larger models running in data centres and reached over the network. Apple has bet that a great deal of useful AI can and should run locally, with the cloud reserved for the harder cases and held to the same privacy standard. If that bet pays off, it points to a future in which much everyday AI is quiet, private and close to the user rather than remote and metered.

What to watch is how far on-device capability can be pushed and how the balance between local and cloud processing settles as the models improve. Apple typically reveals its plans at its developer conference each year, so the specifics of any given generation are best confirmed against Apple’s own announcements rather than assumed, and readers should check the company’s material for the current details. For broader context on how Apple’s approach compares with the cloud-first labs, our large language models hub tracks the field, and the wider AI section follows the strategies behind it.