Apple Foundation Models

You're viewing the source. Log in to edit.

<style>
.ios-ib th.ios-lab { background-color: #333; color: #fff; text-align: center; width: 28%; padding: 6px; }
.ios-ib td.ios-val { text-align: left; padding: 6px 10px; }
</style>
{| class="ios-ib" style="width: 100%; border: 2px solid #333; text-align: center;" darkstyle="border-color: #48484a;"
|-
| colspan="2" style="padding: 28px 20px 42px;" |
<div style="font: bold 2em/1.2 -apple-system, BlinkMacSystemFont, 'SF Pro Display', 'Helvetica Neue', Helvetica, Arial, sans-serif; color: var(--wiki-color);">Apple Foundation Models</div>
|-
! class="ios-lab" | Developer
| class="ios-val" | [[Apple]]
|-
! class="ios-lab" | Introduced
| class="ios-val" | 2024 (with [[Apple Intelligence]]) · framework at WWDC25
|-
! class="ios-lab" | Models
| class="ios-val" | ~3B-parameter on-device model · larger server model on Private Cloud Compute
|-
! class="ios-lab" | Developer access
| class="ios-val" | Foundation Models framework (Swift) — free on-device inference
|}

{{Apple Navbox}}

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.

= 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.

= 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.

Loading editor…