The models behind Apple Intelligence
Apple Foundation Models are the family of models Apple develops to power Apple Intelligence. They handle language understanding and generation, image understanding, and tool calling, running either on the device or on Apple’s servers depending on the task.
It is not one enormous model but a family: lightweight models that run on iPhone and Mac, models in Private Cloud Compute for heavier work, and specialized models for tasks such as image generation.
When it appeared, and what changed
Apple’s own on-device and server models were first described in detail at WWDC24, alongside Apple Intelligence, so the foundation models themselves are not new in WWDC26. WWDC25 brought the second generation and the Foundation Models framework that lets developers use them.
WWDC26 introduced the third generation, and the name became the central way of describing Apple Intelligence’s architecture. The new generation was developed in collaboration with Google, drawing on Gemini model technology, and runs both on device and in Private Cloud Compute.
On-device and cloud
- On-device models: fast to respond, and able to process input without it leaving the device
- Private Cloud Compute models: for questions and reasoning that need more computation
Choosing where processing happens according to the difficulty of the task, rather than sending everything to the cloud, is the defining characteristic of Apple Intelligence.
How it differs from the Foundation Models framework
Apple Foundation Models are the models. The Foundation Models framework is the developer API for using them from an app — the difference between the reasoning engine and the mechanism that connects it to an application.
Generations, supported languages, and available features change with OS updates. When reading about a feature, check which generation and runtime it refers to.