Core ML Reference

A lookup reference for Core ML's Swift API, checked against the iOS 27.1 SDK, plus an index of the coremltools Python APIs for conversion, compression, and multifunction models. It's for developers who know what they want to do and need the exact API, its availability, or whether a type can cross concurrency boundaries.

When to Use This Reference

Use this reference when you need:

• The exact signature or availability of an MLModel, MLModelAsset, or MLModelConfiguration API
• Which compute devices a device offers, and how to check them in code
• Which prediction overload to call, and which Core ML types are Sendable under Swift 6
• MLState and MLTensor operations
• MLComputePlan for per-operation device placement
• Which deployment target an API or coremltools feature requires
• The right coremltools function or config class for a conversion or compression step

Example Prompts

Questions you can ask your assistant that will draw from this reference:

• "What's the signature of MLModel.load, and which URL does it expect?"
• "Is MLModel Sendable? What about MLTensor and MLState?"
• "How do I get the top-k values from an MLTensor?"
• "How do I read which device each operation runs on with MLComputePlan?"
• "Which coremltools config class does 4-bit per-block quantization?"
• "What do I need for a multifunction model, and which iOS version?"

What's Covered

Model Lifecycle

MLModel.load, MLModel.compileModel(at:), .mlpackage / .mlmodelc, MLModelAsset(url:), MLModelAsset(specification:), modelDescription(of:), MLModelConfiguration, computeUnits, functionName, modelDisplayName, optimizationHints, reshapeFrequency, specializationStrategy

Compute Availability

MLModel.availableComputeDevices, MLComputeDevice, MLCPUComputeDevice, MLGPUComputeDevice, MLNeuralEngineComputeDevice, totalCoreCount, preferredMetalDevice

Prediction

prediction(from:) (sync and async), prediction(from:options:), predictions(from:options:), MLPredictionOptions, outputBackings, cancellation, Sendability of MLModel, MLFeatureProvider, MLState, MLTensor, MLShapedArray, MLSendableFeatureValue

State

MLState, makeState(), withMultiArray(for:_:), stateDescriptionsByName

MLTensor

Creating tensors, shape and scalar type, math (softmax(alongAxis:), squareRoot, pointwiseMin, pointwiseMax, topK), reshaping (reshaped(to:), transposed(permutation:), squeezingShape(), tiled(multiples:), split), indexing, replacing(with:where:), resized(to:method:), materializing with shapedArray(of:), withMLTensorComputePolicy(_:_:), MLComputePolicy

coremltools Conversion

ct.convert, minimum_deployment_target, compute_precision, input and output types, flexible shapes, ct.StateType

coremltools Compression

coremltools.optimize.coreml (OpLinearQuantizerConfig, OpPalettizerConfig, OpThresholdPrunerConfig, per_grouped_channel, per_block), coremltools.optimize.torch (PostTrainingQuantizer, PostTrainingPalettizer, SKMPalettizer, DKMPalettizer, LayerwiseCompressor, SparseGPT)

Multifunction Models

ct.utils.MultiFunctionDescriptor, add_function, default_function_name, ct.utils.save_multifunction, functionName

Profiling

MLComputePlan, MLModelStructure, deviceUsage(for:), estimatedCost(of:), the Xcode Core ML performance report, the Core ML instrument

Deployment Targets and Conversion Passes

The per-OS matrix of Core ML APIs and coremltools features, and conversion pass pipelines

Documentation Scope

This page documents the coreml-ref skill (in the axiom-ai suite). Your assistant uses it to look up exact APIs. Its sibling skills cover when and how to use them: deployment patterns, conversion, compression, and diagnostics.

Related

• coreml-deployment – How to use these APIs in an app under Swift 6
• coreml-conversion – Conversion decisions behind the coremltools APIs
• coreml-compression – Compression decisions behind the coremltools.optimize APIs
• coreml-diag – Troubleshooting load failures, slowness, and accuracy loss
• core-ai – The iOS 27 runtime for language-model-scale models

Resources

WWDC: 2023-10049, 2024-10161, 2024-10159

Docs: /coreml/mlmodel, /coreml/mlmodelconfiguration, /coreml/mlmodelasset, /coreml/mlstate, /coreml/mltensor, /coreml/mlcomputedevice, /coreml/mlsendablefeaturevalue, /coreml/mlpredictionoptions