Core ML · Private

On-device AI, where the data never leaves.

Inference on the phone: no round trip, no server bill per request, and no privacy policy explaining where your users' data went.

Stack

Core MLApple IntelligenceVisionCreate ML

01Scope

What on-device intelligence covers.

Apple's frameworks do more than most teams realise before they check.

Apple Intelligence and Foundation Models
The on-device model available through Apple's frameworks, for summarisation, classification and structured generation at no per-token cost.
Core ML model integration
Converting, quantising and profiling models so they fit the memory and thermal budget of the devices you actually support.
Vision and natural language
Text recognition, document scanning, image classification and entity extraction, running offline and instantly.
Hybrid routing
On-device for the common path, server for the hard path, with an explicit rule for which runs when.

02Why it wins

The three arguments for on-device.

Any one of them is usually enough to justify the approach.

Privacy as a feature
Health, finance, journaling and legal products can state that data never leaves the device — and pass the review questions that follow.
Zero marginal cost
No per-request bill. A feature used a hundred times a day by every user does not change your unit economics.
Offline and instant
No network dependency and no round-trip latency, which is the difference between a feature people use and one they wait for.

03FAQ

On-device AI: common questions.

Which devices support on-device AI?

Core ML runs across the supported iPhone range, with performance scaling by chip generation. Apple Intelligence features require the newer devices specifically, so a product depending on them needs a stated minimum device and a fallback for everything below it.

On-device or cloud AI?

On-device when the data is sensitive, the feature runs constantly, or it must work offline. Cloud when the task needs a frontier model's reasoning. Most products are best served by both, with an explicit routing rule.

Can you convert our existing model to Core ML?

Usually — conversion, quantisation and on-device profiling against a real device set. We check size, memory, thermals and accuracy loss before committing to it in a build.

Tell us what you're building.

Send the rough version. A senior engineer reads every brief and replies within one business day.