Google ships ML Kit only through CocoaPods. This project rebuilds those pods as XCFrameworks, publishes them as GitHub Release assets, and exposes every user-facing ML Kit iOS API as a SwiftPM library product.
- iOS 15.5 or later (the minimum the ML Kit 9 pods declare)
- Xcode 15 or later to consume the package; Xcode with Swift 6.4 to build it
Apple Silicon simulators are supported — see Simulator support.
dependencies: [
.package(url: "https://github.com/d-date/google-mlkit-swiftpm", from: "9.0.4")
]Then add the modules you need:
.target(
name: "YourTarget",
dependencies: [
.product(name: "MLKitBarcodeScanning", package: "google-mlkit-swiftpm"),
.product(name: "MLKitTextRecognition", package: "google-mlkit-swiftpm"),
]
)Three steps are not optional. Skip any of them and the module fails at runtime rather than at build time:
- Add the linker flags
- Add the resource bundles your modules need
- Nothing else — but read Simulator support if you are
coming from an older release with
EXCLUDED_ARCHSset.
Upgrading to 9.0.4? It is built with swiftpods instead of the old Makefile pipeline and raises the minimum to iOS 15.5. Resolving it prints two
#warnings; see Dependency overrides.Upgrading from 9.0.0 or 9.0.0-1?
9.0.3is the same upstream ML Kit 9.0.0 build, repackaged. Since 9.0.1, text recognition, pose detection and selfie segmentation ship their models at all, a single product can be adopted without undefined symbols, the simulator slice covers arm64, and App Store Connect accepts the upload. If you setEXCLUDED_ARCHS[sdk=iphonesimulator*] = arm64for an earlier release, remove it.
Add these to Other Linker Flags for your app target:
-ObjC-all_load
ML Kit ships as static archives that register classes and models through
Objective-C categories and static initialisers. Without these flags the linker
drops the archive members that hold them, and you get unrecognized selector
or a model that refuses to load.
Every ML Kit model lives in a resource bundle, and Swift Package Manager cannot carry a resource bundle inside a binary target. Each bundle is published as a separate release asset: download the ones your modules need, add them to your Xcode project, and make sure they land in Copy Bundle Resources of your app target.
./scripts/download_bundles.sh 9.0.4| Bundle | Needed by |
|---|---|
GoogleMVFaceDetectorResources.bundle |
MLKitFaceDetection |
LatinOCRResources.bundle |
MLKitTextRecognition |
ChineseOCRResources.bundle |
MLKitTextRecognitionChinese |
DevanagariOCRResources.bundle |
MLKitTextRecognitionDevanagari |
JapaneseOCRResources.bundle |
MLKitTextRecognitionJapanese |
KoreanOCRResources.bundle |
MLKitTextRecognitionKorean |
MLKitImageLabelingResources.bundle |
MLKitImageLabeling |
MLKitObjectDetectionResources.bundle |
MLKitObjectDetection |
MLKitObjectDetectionCommonResources.bundle |
MLKitImageLabeling, MLKitImageLabelingCustom, MLKitObjectDetection, MLKitObjectDetectionCustom |
MLKitPoseDetectionFastResources.bundle |
MLKitPoseDetection |
MLKitPoseDetectionAccurateResources.bundle |
MLKitPoseDetectionAccurate |
MLKitPoseDetectionCommonResources.bundle |
MLKitPoseDetection, MLKitPoseDetectionAccurate |
MLKitSegmentationSelfieResources.bundle |
MLKitSegmentationSelfie |
MLKitSegmentationCommonResources.bundle |
MLKitSegmentationSelfie |
MLKitXenoResources.bundle |
MLKitPoseDetection, MLKitPoseDetectionAccurate, MLKitSegmentationSelfie |
MLKitTranslate_resource.bundle |
MLKitTranslate |
MLKitDigitalInkRecognition_resource.bundle |
MLKitDigitalInkRecognition |
PredictOnDeviceResource.bundle |
MLKitSmartReply |
PredictOnDevice_resource.bundle is also published: it is the same content
under the name ML Kit nests inside its own framework, which is what releases
before 9.0.1 shipped. Add whichever your build already refers to; adding both
is harmless.
Since 9.0.4 the bundles of the two pods built from source are published too:
GoogleToolboxForMac_Privacy.bundle, GoogleToolboxForMac_Logger_Privacy.bundle
and SSZipArchive.bundle.
A missing bundle shows up as a runtime throw — text recognition, for example,
raises MLKTextRecognizerInternalErrorCreationFailure with "Invalid model
path.".
Add only the products you use; each one links just the frameworks it needs.
| Product | What it does |
|---|---|
MLKitBarcodeScanning |
Scan and decode barcodes |
MLKitFaceDetection |
Detect faces, contours and facial features |
MLKitTextRecognition |
Recognise Latin-script text (v2) |
MLKitTextRecognitionChinese |
Chinese text |
MLKitTextRecognitionDevanagari |
Devanagari text |
MLKitTextRecognitionJapanese |
Japanese text |
MLKitTextRecognitionKorean |
Korean text |
MLKitImageLabeling |
Label objects, places and activities |
MLKitImageLabelingCustom |
Image labeling with your own model |
MLKitObjectDetection |
Detect and track objects |
MLKitObjectDetectionCustom |
Object detection with your own model |
MLKitPoseDetection |
Body pose detection |
MLKitPoseDetectionAccurate |
Body pose detection, accurate model |
MLKitSegmentationSelfie |
Separate people from the background |
| Product | What it does |
|---|---|
MLKitLanguageID |
Identify the language of a string |
MLKitTranslate |
Translate between languages on device |
MLKitSmartReply |
Suggest contextual replies |
MLKitEntityExtraction |
Find addresses, dates, phone numbers and the like in text |
MLKitDigitalInkRecognition |
Recognise handwriting and drawn strokes |
ML Kit's pre-built binaries contain no arm64 simulator code, and Xcode 26 dropped Rosetta simulators — so on Apple Silicon the published XCFrameworks used to be unusable in the Simulator entirely.
Since 9.0.1 the arm64 simulator slice is synthesised from the device slice by
rewriting the Mach-O platform (done by swiftpods package since 9.0.4).
Published XCFrameworks carry arm64 for iphoneos and arm64 + x86_64 for
iphonesimulator, so no EXCLUDED_ARCHS workaround is needed.
Two Google dependencies are declared wider than the ML Kit pods accept, in
Podfile.swift with overridingPodRequirements: .unsafe(reason:), because
current Firebase needs them:
gtm-session-fetcher3.5.0..<6.0.0(the pods require>= 3.3.2, < 4.0); the ABI was checked by hand for Firebase 13 (#124, #125).nanopb2.30910.0..<2.30911.0(the pods require~> 3.30910.0); Firebase 13.1.1 requires2.30910.1, a fork commit of the pods'2.30910.0.
The generated Package.swift starts with one #warning per override, which
swift package resolve prints when it evaluates the manifest. It is expected
and does not need action.
- No dSYMs. Google ships ML Kit without debug info, so Xcode's archive step reports "Upload Symbols Failed" for each ML Kit framework. The frameworks are linked statically, so their symbols end up in your app's own dSYM; the warnings are spurious and do not block submission. You could not symbolicate inside ML Kit anyway — it is closed source.
- Resource bundles are manual. Swift Package Manager has no way to ship them inside a binary target. See above.
- Digital ink recognition and entity extraction download their models at runtime from Google, so they ship no bundled model and need network access the first time you use them.
- The wrapper version is not the pod version. Major and minor follow ML
Kit; the patch number counts wrapper releases.
9.0.4repackages upstream ML Kit9.0.0; the pod versions each framework reports are Google's own.
./scripts/download_bundles.sh 9.0.4
cd Example && open Example.xcworkspaceSet code signing to your own team. The app runs on a device and on the Apple Silicon Simulator, and exercises every module.
The package is built by swiftpods from
Podfile.swift. swiftpods package builds the pods, writes every
*.xcframework.zip and *.bundle.zip to build/distribution/, regenerates
Package.swift (do not edit it by hand) and links every product on its own.
CLAUDE.md and AUTOMATION.md have the details.
swiftpods package --version <version> # build into build/distribution/, regenerate Package.swift
./scripts/verify_local_archive.sh # run ML Kit on the Simulator, archive for device
./scripts/verify_app_store_upload.sh --upload # Apple-side validation and delivery
scripts/release.sh <version> # create the GitHub Release and upload the zipsverify_local_archive.sh is the gate that matters: it runs inference on the
arm64 Simulator, launches the app, archives for device, and checks that every
model bundle reached the app and that nothing was linked dynamically. A missing
model bundle builds and archives perfectly and only fails when inference runs,
so build-time checks alone are not enough.
To pick up a new upstream ML Kit version, raise the GoogleMLKit/*
requirements in Podfile.swift, run swiftpods update, and release with the
ML Kit major.minor (ruby scripts/check_mlkit_version.rb reports whether one
is out). Normally run the Build MLKit XCFrameworks workflow from the
Actions tab; it packages, commits, tags and calls scripts/release.sh.