Model components / Inference code

Apple Core ML Runtime

Apple Core ML is an on-device inference runtime that executes models across Apple's CPU, GPU, and Neural Engine, shipped as a system framework with iOS, macOS, watchOS, tvOS, and visionOS. It partitions model graphs across the available hardware and runs .mlpackage and .mlmodel files, which the separately distributed coremltools package produces. Apple's own on-device features, including Apple Intelligence, are built on it.

Verified 2026-08-09 via Apple's Core ML documentation, the MLModel availability list, and the coremltools optimization guide.

Openness

1 high confidence

1.0

source: closed
license: proprietary(Apple system framework, ships compiled with the OS/SDK, no published source)
self-host: no

Core ML has no repository. Apple's own documentation describes it purely as an API surface, MLModel and the rest, that an app calls into; there is no source tree to self-host or fork. Whether any feature is held back behind a paid tier does not arise here, because that question only applies where the source is published at all.

https://developer.apple.com/tutorials/data/documentation/coreml.md recorded 2026-08-09
Use Core ML to integrate machine learning models into your app... Copyright © 2026 Apple Inc. All rights reserved.

Adoption

5 medium confidence

5.0

Core ML ships as a system framework across iOS, iPadOS, macOS, tvOS, watchOS and visionOS, confirmed against Apple's MLModel availability list, and Apple's own quarterly earnings call puts the active-device installed base at more than 2.5 billion. That bands at >10M users, level 5. The quantity being banded is an active device base rather than an active user count: someone with an iPhone and a Mac counts twice, and nobody is counted for using Core ML rather than for owning hardware it ships on. It is not a developer download count either. The figure clears the band boundary by more than two orders of magnitude, so the substitution does not change the level, but the scale requires a substituted quantity to be named.

https://docs.developer.apple.com/tutorials/data/documentation/coreml/mlmodel.md recorded 2026-08-09
"availability" : [ "iOS: 11.0.0 -", "iPadOS: 11.0.0 -", "macCatalyst: 13.1.0 -", "macOS: 10.13.0 -", "tvOS: 11.0.0 -", "visionOS: 1.0.0 -", "watchOS: 4.0.0 -" ]
https://sixcolors.com/post/2026/07/one-last-time-this-is-tim-transcript-of-apples-q3-2026-financial-call/ recorded 2026-08-09
installed base of over two and a half billion active devices has reached another all-time high across all major product categories and geographic segments

Capability

4 medium confidence

4.0

Best in class for on-device and edge inference, with Neural Engine acceleration, quantization and palettization, and broad coverage of converted models, but constrained to Apple silicon and to on-device scale. That places it one tier below the large-scale server engines, of which vLLM holds the top capability score in this category.

https://docs.developer.apple.com/tutorials/data/documentation/coreml.md recorded 2026-08-09
leveraging the CPU, GPU, and Neural Engine while minimizing its memory footprint and power consumption
https://apple.github.io/coremltools/docs-guides/source/opt-overview.html recorded 2026-08-09
Core ML supports INT4 and INT8 quantization options for weights and INT8 for activations.

Verified 2026-08-09