CreateML

This Create ML review covers what Apple's model trainer does, its key features, what it costs, how the Create ML app differs from the framework, how it relates to Core ML, where it shines and falls short, and how it compares with other machine learning tools. Create ML trains custom machine learning models on your own computer from your own training data. The process is short: you pick a template, drop in labeled examples, press Train, check the validation accuracy and export a Core ML model file that Xcode turns into a Swift class. You never pick algorithms or write training loops; Create ML chooses a model structure that fits the task. Apple builds it on the machine learning infrastructure behind Photos and Siri, so image and language models stay small and train quickly (Apple Developer Documentation).

What is Create ML used for?

Create ML is used by iOS, iPadOS, macOS, watchOS and visionOS developers to add on-device predictions to their apps without writing training code. Typical examples are recognizing images of products or plants, object detection with bounding boxes, classifying support messages or reviews by text, tagging words in a sentence, recognizing sounds, detecting body or hand actions in video, classifying motion from Apple Watch sensors, and predicting values from tabular data. A recipe app can recognize dishes from a photo, a fitness app can count exercises from video, and a note app can sort text by topic. Text models can also do sentiment analysis in several languages. Because the model runs on Apple devices through Core ML, predictions work offline, respond at native speed and user data stays on the phone.

Key features

• Templates for each task: image classification, object detection, hand pose, action and hand action classification, style transfer, text classification, word tagging, sound classification, activity classification, tabular classification, tabular regression and recommendation.
• Spatial models: an object tracking template trains reference objects for visionOS apps on Apple Vision Pro, added in 2024 (WWDC24).
• Transfer learning: most templates start from Apple's pretrained feature extractors, so a useful model needs far less data than training from scratch. Apple asks for at least 10 images per category for an image classifier, and more varied images work better (Apple).
• Data augmentation: toggles for rotation, flip, blur, crop and exposure create extra training images on the fly. Augmentation can raise training accuracy while lowering validation accuracy, so compare both.
• Evaluation and live preview: training and validation accuracy, a confusion matrix that shows correct and wrong predictions per class, root mean squared error for regression, and a preview tab where you drop in new data to see predictions before export. A large gap between training and validation accuracy points to overfitting.
• Training control: pause, resume and compare several models in one project; the framework adds checkpoints and training sessions.
• Model metadata: author name, description, version and license are saved into the exported model.
• Swift framework: the same models can be trained in code with the Create ML framework, for scripts, automation and training inside an app.

Create ML pricing

Create ML is free. The Create ML app ships with Xcode, which is a free download from the Mac App Store, and the framework is part of Apple's SDKs. There is no subscription, usage metric or cloud bill: training runs on your own Mac, so the cost is your hardware and time. To put a trained model into an App Store app you need an Apple Developer Program membership, which Apple charges yearly for publishing, not for Create ML itself. The cost surprise to check is hardware: large image or video datasets train much faster on Apple silicon than on an older Intel Mac, and there is no option to rent cloud GPUs from inside Create ML.

Create ML app vs the Create ML framework

The Create ML app is the no-code option. You open it from Xcode with Open Developer Tool > Create ML, choose a template, drag in folders of training data and press the Train button (Hacking with Swift). To import data, image models use one subfolder per class, text classifiers take folders or a flat JSON array, object detection needs a JSON file with a bounding box per object, and tabular models read CSV or JSON, where the model learns relationships between columns. If you add no validation data, the Create ML app holds back about 20% of the training data automatically (Netguru). When training completes, it shows the metrics and an output tab, where you save the mlmodel file.

The Create ML framework is the code option. In a Swift playground or macOS command-line tool you create an instance such as MLImageClassifier or MLRegressor, pass an MLDataTable or a folder, and read the metrics from code (Apple Developer Documentation). The framework suits repeatable pipelines and continuous integration; the Create ML app suits exploring a dataset, comparing models by eye and quick one-off training. Both export the same Core ML model file. Apple lists the framework on macOS 10.14 and later and, for some model types, on iOS and iPadOS 15 and later, so an app can train or update a model on the device. Recent templates need a current Xcode on a recent macOS; one guide puts the floor at macOS Ventura and Xcode 14 (Netguru).

Create ML vs Core ML

Create ML and Core ML do different jobs. Create ML trains the model; Core ML runs it inside your app (Core ML documentation). When you add the exported file to an Xcode project, Xcode generates a Swift class with typed inputs and outputs, and Core ML decides whether to run the Core ML model on the CPU, GPU or Neural Engine for the best performance on each device. Models trained elsewhere, for example in PyTorch or TensorFlow, reach Core ML through the open-source Core ML Tools, which convert models to the Core ML format. So you can start with Create ML for a first version and move to your own training code later without changing how the app calls the model.

Where Create ML shines — and where it falls short

Strengths:

• Free, built into Xcode and fast to learn for Swift developers.
• Small models that run offline and keep data on the device.
• Good results with little data thanks to transfer learning.
• Covers images, video, text, sound, motion, tabular data and spatial objects.

Trade-offs:

• Mac only, and models target Apple platforms through Core ML.
• No control over model architecture beyond a few parameters.
• No cloud training, team features, feedback loops or experiment tracking beyond one project.
• Not built for large language models or generative AI.

How Create ML compares

• Teachable Machine: Google's free browser tool for quick image, sound and pose models, exported for the web and TensorFlow.
• Obviously AI: no-code predictions on business data in the cloud, aimed at analysts rather than app developers.
• DataRobot: enterprise AutoML and model operations for data science teams.
• H2O.ai: open-source and enterprise AutoML for tabular data at scale.
• Dataiku: a collaborative platform for data preparation, machine learning and deployment across a company.

Is Create ML worth it?

Create ML is worth it for Apple developers who want a custom image, text, sound or tabular model inside their app without learning a training framework. It is free, quick to try and produces models that run fast and privately on the device. It is less useful for teams that need cross-platform models, cloud training, full control over the network or generative AI; those teams use PyTorch or TensorFlow and convert the result for Core ML. If you want help building a model or adding machine learning features to an iOS app, the developers on LowCodeDevs can take it on.

Sources

• Create ML, Apple Developer Documentation: model types, framework classes and platforms.
• Core ML, Apple Developer Documentation: running models in apps.
• Introducing the Create ML App, WWDC19: the app workflow.
• What's new in Create ML, WWDC24: object tracking for visionOS and time-series APIs.
• Training a model with Create ML, Hacking with Swift: a step-by-step tabular regression example.
• Create ML guide, Netguru: app vs framework, data formats and augmentation.