Anhang

📚Quellen & Stand

Alle Versions- und Fachangaben sind gegen diese Seiten geprüft (Stand 28.09.2026). Apple-Seiten wurden über die Daten der Developer Documentation gelesen (inkl. „Available since“-Angaben).

🏷️Versionsstände

coremltools (stabil)9.0 vom 10.11.2025
coremltools (Vorabversion)9.1.dev1 vom 03.08.2026
getestetes PyTorch2.7.0 – „Torch 2.7.0 is the most recent version that has been tested.“ (Warnung von coremltools 9.0 beim Import mit torch 2.7.1)
Deployment-ZieleiOS13 … iOS18, iOS26 (ct.target in coremltools 9.0)
PythonWheels für Python 3.7 – 3.13 (9.0 auf PyPI)
Eigene Konvertierungcoremltools 9.0, torch 2.7.0, numpy 2.2.6, Python 3.12.14, macOS 27.0

🍏Apple

Core ML (Framework-Übersicht)
Vision, Natural Language, Speech, Sound Analysis bauen auf Core ML auf; CPU, GPU, Neural Engine; ab iOS 11 / macOS 10.13
https://developer.apple.com/documentation/coreml
MLModel
generierte Wrapper-Klasse, Thread-Hinweis, compileModel, predictions(fromBatch:)
https://developer.apple.com/documentation/coreml/mlmodel
MLModelConfiguration
ab iOS 12 / macOS 10.14; computeUnits, allowLowPrecisionAccumulationOnGPU, optimizationHints
https://developer.apple.com/documentation/coreml/mlmodelconfiguration
MLComputeUnits
.all/.cpuOnly/.cpuAndGPU ab iOS 12 / macOS 10.14; .cpuAndNeuralEngine ab iOS 16 / macOS 13
https://developer.apple.com/documentation/coreml/mlcomputeunits
MLComputeUnits.cpuAndNeuralEngine
https://developer.apple.com/documentation/coreml/mlcomputeunits/cpuandneuralengine
MLModel.compileModel(at:) (synchron)
ab iOS 11 / macOS 10.13; in der aktuellen Doku mit Deprecation-Vermerk 27.2 geführt
https://developer.apple.com/documentation/coreml/mlmodel/compilemodel(at:)-6442s
MLModel.compileModel(at:) async
ab iOS 16 / macOS 13
https://developer.apple.com/documentation/coreml/mlmodel/compilemodel(at:)-45ao6
MLModel.load(contentsOf:configuration:) async
ab iOS 15 / macOS 12
https://developer.apple.com/documentation/coreml/mlmodel/load(contentsof:configuration:)
MLModel predictionFromFeatures:completionHandler: (in Swift: async prediction(from:))
ab iOS 17 / macOS 14
https://developer.apple.com/documentation/coreml/mlmodel/predictionfromfeatures:completionhandler:
MLModel.predictions(fromBatch:)
ab iOS 12 / macOS 10.14
https://developer.apple.com/documentation/coreml/mlmodel/predictions(frombatch:)
MLMultiArray
ab iOS 11 / macOS 10.13
https://developer.apple.com/documentation/coreml/mlmultiarray
MLShapedArray
ab iOS 15 / macOS 12
https://developer.apple.com/documentation/coreml/mlshapedarray
MLTensor
ab iOS 18 / macOS 15
https://developer.apple.com/documentation/coreml/mltensor
MLState
ab iOS 18 / macOS 15
https://developer.apple.com/documentation/coreml/mlstate
MLUpdateTask
ab iOS 13 / macOS 10.15, nicht als veraltet markiert
https://developer.apple.com/documentation/coreml/mlupdatetask
Personalizing a Model with On-Device Updates
Update-Task nur für kompilierte .mlmodelc
https://developer.apple.com/documentation/coreml/personalizing-a-model-with-on-device-updates
MLComputePlan
ab iOS 17.4 / macOS 14.4
https://developer.apple.com/documentation/coreml/mlcomputeplan-1w21n
Downloading and Compiling a Model on the User’s Device
https://developer.apple.com/documentation/coreml/downloading-and-compiling-a-model-on-the-user-s-device
Updating a Model File to a Model Package
https://developer.apple.com/documentation/coreml/updating-a-model-file-to-a-model-package
Integrating a Core ML Model into Your App
Xcode erzeugt Model-, Input- und Output-Klassen
https://developer.apple.com/documentation/coreml/integrating-a-core-ml-model-into-your-app
Analyzing a Core ML model’s performance in Xcode
https://developer.apple.com/documentation/coreml/analyzing-a-core-ml-model-s-performance-in-xcode
Encrypting a Model in Your App
https://developer.apple.com/documentation/coreml/encrypting-a-model-in-your-app
MLModelCollection (Core ML Model Deployment)
ab iOS 14, veraltet seit iOS 17.4
https://developer.apple.com/documentation/coreml/mlmodelcollection
Getting a Core ML Model
https://developer.apple.com/documentation/coreml/getting-a-core-ml-model
Core ML Models (Apple Machine Learning)
u. a. FastViT, MobileNetV2, ResNet-50, Depth Anything V2, DETR, DeepLabV3, MNIST, BERT-SQuAD, YOLOv3
https://developer.apple.com/machine-learning/models/
Create ML
https://developer.apple.com/documentation/createml
VNCoreMLModel
ab iOS 11 / macOS 10.13
https://developer.apple.com/documentation/vision/vncoremlmodel
VNCoreMLRequest
Ergebnistyp abhängig vom Modell; Konfidenzen werden nicht normalisiert
https://developer.apple.com/documentation/vision/vncoremlrequest
VNClassificationObservation
https://developer.apple.com/documentation/vision/vnclassificationobservation
CoreMLRequest (Swift-Vision-API)
ab iOS 18 / macOS 15; perform(on:) async
https://developer.apple.com/documentation/vision/coremlrequest
CoreMLModelContainer
init(model:featureProvider:) ab iOS 18 / macOS 15
https://developer.apple.com/documentation/vision/coremlmodelcontainer
ClassificationObservation
identifier, confidence; ab iOS 18 / macOS 15
https://developer.apple.com/documentation/vision/classificationobservation
Core AI (neues Framework)
ab iOS 27 / macOS 27; Format .aimodel, Konverter coreai-torch; für andere Modelltypen verweist Apple auf Core ML
https://developer.apple.com/documentation/coreai

🧰coremltools

coremltools Release Notes (GitHub)
9.0 vom 10.11.2025: Python 3.13, Ziele iOS26/macOS26, PyTorch 2.7, int8-Ein-/Ausgänge
https://github.com/apple/coremltools/releases
coremltools 9.0 Release Notes
https://github.com/apple/coremltools/releases/tag/9.0
coremltools 8.3 Release Notes
Debugging- und Performance-Werkzeuge: MLModelValidator, MLModelComparator, MLModelInspector, MLModelBenchmarker
https://github.com/apple/coremltools/releases/tag/8.3
coremltools 8.0 Release Notes
torch.export bei 56 % Parität, 4-Bit-Quantisierung, Stateful Models
https://github.com/apple/coremltools/releases/tag/8.0
coremltools 6.0 Release Notes
„Remove ONNX support.“
https://github.com/apple/coremltools/releases/tag/6.0
coremltools auf PyPI
9.0 = neueste stabile Version; 9.1.dev1 (03.08.2026) Vorabversion
https://pypi.org/project/coremltools/
What Is Core ML Tools?
unterstützte Bibliotheken: TensorFlow, PyTorch, XGBoost, scikit-learn, LIBSVM
https://apple.github.io/coremltools/docs-guides/source/overview-coremltools.html
PyTorch Conversion Workflow
trace = stabiler Weg, torch.export seit 8.0 (Beta); eval() vor dem Tracen
https://apple.github.io/coremltools/docs-guides/source/convert-pytorch-workflow.html
Source and Conversion Formats
ab 7.0 Standard mlprogram (iOS15/macOS12); neuralnetwork im Wartungsmodus
https://apple.github.io/coremltools/docs-guides/source/target-conversion-formats.html
Convert Models to ML Programs
FP16 Standard, compute_precision, nur .mlpackage
https://apple.github.io/coremltools/docs-guides/source/convert-to-ml-program.html
Comparing ML Programs and Neural Networks
Gewichte getrennt; model.mil im .mlmodelc; On-Device-Update nur NeuralNetwork
https://apple.github.io/coremltools/docs-guides/source/comparing-ml-programs-and-neural-networks.html
Image Input and Output
y = x·scale + bias; torchvision scale = 1/(0.226·255)
https://apple.github.io/coremltools/docs-guides/source/image-inputs.html
Flexible Input Shapes
EnumeratedShapes bis 128 Formen; unbeschränkte Bereiche bei mlprogram nicht erlaubt
https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html
Classifiers
https://apple.github.io/coremltools/docs-guides/source/classifiers.html
Model Prediction
predict() nur unter macOS
https://apple.github.io/coremltools/docs-guides/source/model-prediction.html
MLModel Overview (Metadaten, Spec, predict)
model.author, short_description, input_description …
https://apple.github.io/coremltools/docs-guides/source/mlmodel.html
MLModel Utilities (Metadaten, rename_feature)
https://apple.github.io/coremltools/docs-guides/source/mlmodel-utilities.html
Composite Operators
@register_torch_op
https://apple.github.io/coremltools/docs-guides/source/composite-operators.html
Debugging And Performance Utilities
MLModelInspector/Validator/Comparator – experimentell
https://apple.github.io/coremltools/docs-guides/source/mlmodel-debugging-perf-utilities.html
FAQs
onnx-coreml eingefroren
https://apple.github.io/coremltools/docs-guides/source/faqs.html
Optimization Overview
Palettisierung 1,2,3,4,6,8 Bit; INT4/INT8-Gewichte; INT8-Aktivierungen
https://apple.github.io/coremltools/docs-guides/source/opt-overview.html
Optimization – What’s New (OS-Verfügbarkeit)
https://apple.github.io/coremltools/docs-guides/source/opt-whats-new.html
Quantization Overview + API
symmetrisch ist Standard; W8A8 auf NE ab A17 Pro/M4
https://apple.github.io/coremltools/docs-guides/source/opt-quantization-overview.html
Quantization Performance
per_block eher GPU, auf der NE per_channel empfohlen
https://apple.github.io/coremltools/docs-guides/source/opt-quantization-perf.html
Palettization Overview + API
https://apple.github.io/coremltools/docs-guides/source/opt-palettization-overview.html
Palettization Performance
https://apple.github.io/coremltools/docs-guides/source/opt-palettization-perf.html
Pruning Overview + API
https://apple.github.io/coremltools/docs-guides/source/opt-pruning-overview.html
Pruning Performance
https://apple.github.io/coremltools/docs-guides/source/opt-pruning-perf.html
scikit-learn conversion
https://apple.github.io/coremltools/docs-guides/source/sci-kit-learn-conversion.html
XGBoost conversion
https://apple.github.io/coremltools/docs-guides/source/xgboost-conversion.html
Converting from TensorFlow 2
https://apple.github.io/coremltools/docs-guides/source/tensorflow-2.html
Stateful Models
https://apple.github.io/coremltools/docs-guides/source/stateful-models.html
API-Referenz: coremltools.convert
https://apple.github.io/coremltools/source/coremltools.converters.convert.html
API-Referenz: TensorType, ImageType, RangeDim, EnumeratedShapes, ClassifierConfig
https://apple.github.io/coremltools/source/coremltools.converters.mil.input_types.html
API-Referenz: coremltools.optimize.coreml
https://apple.github.io/coremltools/source/coremltools.optimize.coreml.post_training_quantization.html
API-Referenz: MIL Ops
https://apple.github.io/coremltools/source/coremltools.converters.mil.mil.ops.defs.html

🔥PyTorch

torch.export (Doku PyTorch 2.7)
„prototype under active development“ (bis 2.8)
https://docs.pytorch.org/docs/2.7/export.html
torch.export (aktuelle Doku)
ExportedProgram, dynamic_shapes mit torch.export.Dim
https://docs.pytorch.org/docs/stable/export.html
TorchScript / torch.jit (aktuelle Doku)
seit PyTorch 2.9: „TorchScript is deprecated, please use torch.export instead.“
https://docs.pytorch.org/docs/stable/jit.html
torch.jit.trace (Doku PyTorch 2.7)
https://docs.pytorch.org/docs/2.7/generated/torch.jit.trace.html
pytorch/examples – mnist/main.py
https://github.com/pytorch/examples/blob/main/mnist/main.py
torchvision Modell-Metadaten (num_params)
https://docs.pytorch.org/vision/stable/models.html

🔗Hugging Face

Hugging Face: Organisation apple (Core-ML-Modelle)
z. B. apple/coreml-FastViT-MA36, apple/coreml-depth-anything-small, apple/coreml-stable-diffusion-v1-5
https://huggingface.co/apple

🔗scikit-learn

sklearn.datasets.load_digits
https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_digits.html

🧪Eigene Messung

Eigene Konvertierung des Lehrmodells (coremltools 9.0, torch 2.7.0, macOS 27)
Golden-Files im Repo: tests/golden, Skript scripts/golden/gen_golden.py
https://github.com/apple/coremltools