Loading history…
Apple Foundation Models
This is r2 (the current revision), saved by jack. Full history
|
Apple Foundation Models
| |
| Developer | Apple |
|---|---|
| Introduced | 2024 (with Apple Intelligence) · framework at WWDC25 |
| Models | ~3B-parameter on-device model · larger server model on Private Cloud Compute |
| Developer access | Foundation Models framework (Swift) — free on-device inference |
[ Show / Hide ]
| ||||||||||||||||||||||||||||||||||||||||||||
The Apple Foundation Models are the language models underneath Apple Intelligence: a roughly three-billion-parameter model compressed to run on iPhones and Macs, and a larger sibling running on Apple silicon servers in Private Cloud Compute. Apple documents them in periodic technical reports — training, quantization, adapters, and responsible-AI evaluations — an unusual habit of disclosure for a company that ships models silently inside operating systems.
1 The framework✎
At WWDC 2025 the models became a platform: the Foundation Models framework gives any app direct Swift access to the on-device model — guided generation into typed data structures, tool calling, and streaming — at zero inference cost to the developer and zero data leaving the device. It is Apple's counter-offer to cloud-API economics: a smaller model, but free, private, and always present.
2 Position✎
The AFMs embody Apple's AI position — models sized to the hardware in hand rather than raced for benchmarks, specialized through adapters for the features Apple Intelligence actually ships. They sit between the open research artifacts (OpenELM, MLX) and the product surface, as the working engine of the system.
Content available under CC BY-SA 4.0