Octomil vs Apple Foundation Models

Apple'sFoundation Models frameworkis very good, and it is free. If your app is iOS only and your users are on recent hardware, you should probably use it. This page is about the part of the fleet it does not reach, and the operational layer it does not try to be.

Start from what Apple already gives you

We are not going to pretend the gap is bigger than it is.At WWDC 2026 Apple introduced a provider protocol so aLanguageModelSession can be backed by the on-device system model, Private Cloud Compute, a local MLX model, or a third-party provider, with the session code unchanged. On Apple platforms, "one API, swap the backend" is now a first-party feature. Any comparison written before that is out of date, including older versions of ours.

Use Apple's framework if

  • you ship to Apple platforms only
  • your users are on iOS 26 or later with Apple Intelligence hardware
  • the system model handles your task and you do not need to choose the weights
  • you want zero cost, zero infrastructure, and a Swift-native API

Add Octomil if

  • the same feature also has to ship on Android, the web, or desktop
  • a meaningful share of your installed base cannot run Apple Intelligence
  • you need staged rollout, rollback, and per-cohort quality monitoring of the model
  • you need one policy and one measurement surface across all of it

Capability by capability

CapabilityApple Foundation ModelsOctomil
Cost of on-device inferenceFreeUnmetered on every plan
Quality of the first-party on-device modelStrong, tuned for Apple siliconDepends on the model you choose
Swap model backends behind one APIYes, on Apple platforms, since the 2026 provider protocolYes, and the same policy applies off Apple platforms
Android, web, and desktopNot coveredCovered by the same routing policy
Devices without Apple Intelligence supportNo on-device pathSmaller model or cloud fallback by policy
Staged rollout and rollback by device cohortNot the purpose of the frameworkBuilt in
Fleet quality monitoring after releaseNot providedBuilt in
Cost measured against an all-cloud baselineNot applicableBuilt in

The two questions that decide it

Do you ship anywhere other than Apple?

If yes, you are going to write the local-versus-cloud decision twice, in two languages, against two runtimes, and then keep them consistent as both evolve. That duplicated policy is the thing Octomil removes. If no, this is a much weaker argument and you should take Apple's free path.

What share of your users can actually run it?

The framework needs a recent OS and Apple Intelligence capable hardware. Check that number against your own analytics before assuming the on-device path is available. The remainder still needs a policy, and today that policy is usually an unconditional call to the cloud.

Shipping is the part that stays hard

Neither framework decides when to hold a rollout because one device cohort regressed, or how to roll a model back without shipping a new binary. That work does not disappear because inference became free.

Use Apple's model. Route the rest of the fleet with one policy.

Octomil can call the system model on capable Apple devices and take a different path everywhere else, without forking the decision across platforms.