An honest, developer-focused review of Apple's Create ML — what it does, where it shines, its real limits, and who should actually use it to add machine learning to iOS apps.
Create ML is Apple's tool for training machine learning models with a low-code, mostly no-code interface. It ships as a standalone macOS app alongside Xcode, and as a Swift framework you can call from code.
The pitch is simple. You bring labeled data, pick a template — image, text, sound, tabular, activity, and others — drag your data in, and click Train. Create ML handles the heavy lifting of model architecture and optimization.
What comes out the other end is a Core ML model file. That is the format Apple platforms use for on-device inference across iOS, iPadOS, macOS, watchOS, and tvOS.
It is important to be precise about scope. Create ML trains models. It does not build your app. You still open Xcode, write Swift, wire up the model, and ship through the normal App Store pipeline.
Machine learning on Apple platforms is a two-part story. Create ML is the training half; Core ML is the runtime half.
You train once, typically on your Mac, and produce a model artifact. Then you add that artifact to your Xcode project, and Core ML runs it on the user's device — often accelerated by the Neural Engine, GPU, or CPU depending on the hardware.
This on-device approach is the whole point. Inference happens locally, which means it can work offline, keeps user data on the device, and avoids per-request server costs.
For a typical iOS team, Create ML slots in early. A designer or engineer gathers example data, trains a model in an afternoon, and hands a small file to the app developers. No GPU cluster, no Python environment, and no separate MLOps stack are required for a first version.
The template system is the core feature. Each template targets a common problem — image classification, object detection, text classification, word tagging, sound classification, tabular regression and classification, recommendations, and activity classification, among others.
Transfer learning is a quiet strength. For image and text tasks, Create ML builds on models Apple has already pre-trained, so you often get usable results from a modest number of examples rather than needing millions.
The app gives you a live training view with accuracy and loss curves, plus separate training, validation, and testing splits so you can sanity-check generalization before you ship.
There is also a programmatic path. The Create ML framework lets you script training in Swift, which is handy for reproducible pipelines or retraining as your dataset grows. Data augmentation options for images help squeeze more out of small datasets.
The biggest strength is the time-to-first-model. If you already own a Mac and Xcode, you can go from labeled folders of images to a working classifier in a single sitting, with essentially no ML theory required.
The integration story is genuinely tight. Because output is Core ML, Xcode auto-generates a Swift class for your model, and frameworks like Vision and Natural Language can consume it with very little glue code.
Privacy and offline capability come largely for free. On-device inference means no round-trips to a server, which is a real differentiator for health, camera, and productivity apps.
And the price of entry is low in effort terms. There is no separate cloud account to provision, no data pipeline to stand up, and no infrastructure to babysit just to test an idea.
Create ML is deliberately opinionated, and that is also its ceiling. You get the architectures and knobs Apple exposes, not the full flexibility of a framework like PyTorch or TensorFlow. If you need a custom architecture, you will outgrow it.
It is macOS-only for training. There is no Windows or Linux version, and no first-party cloud training service — training runs on your Mac's hardware, so very large datasets can be slow on lower-end machines.
It does not build, sign, or submit your app. Create ML produces a model; shipping still requires Xcode, code signing, and the Apple Developer Program. No content or training tool changes that.
Debugging is coarse. When accuracy is poor, you get limited insight into why compared to a full ML stack. And the output ecosystem is Apple-centric — a Core ML model is not a drop-in for an Android or web backend without conversion work.
Create ML itself is part of Apple's developer tooling that comes with Xcode and macOS, so there is no separate license fee to train models on your own Mac. Treat that as the general shape rather than a promise — always confirm current terms on Apple's official pages.
The real costs are indirect. You need a Mac capable of running current Xcode, and beefier Apple Silicon will train larger models noticeably faster.
To actually ship an app that uses your model, you need the Apple Developer Program, which is a paid annual membership. That is a requirement for App Store distribution regardless of whether you use machine learning.
For exact, current pricing on the developer program and hardware, check Apple's official pages directly. Avoid budgeting from any number you saw in a blog, including this one.
It helps to place Create ML against the two things developers usually weigh it against.
The first is a full training framework like PyTorch or TensorFlow, followed by a conversion step to Core ML. That route gives you total control over architecture and training, at the cost of a Python environment, more expertise, and more moving parts. Create ML trades that control for speed and simplicity.
The second is a hosted cloud machine learning service. Those can scale to huge datasets and run anywhere, but they add accounts, recurring costs, and network dependencies that undercut the offline, private, on-device experience.
Create ML sits deliberately in the middle-low end of complexity. If your problem fits a template, it is hard to beat on effort. If it does not, one of the alternatives is the honest choice.
Create ML is an easy recommendation for iOS teams that want on-device machine learning without hiring an ML specialist. For image classification, sound recognition, text tagging, and tabular prediction, it is often the fastest path from idea to a shippable Core ML model.
It is ideal for indie developers, small teams, and product engineers who value tight Apple integration, privacy, and offline inference over maximum modeling flexibility.
It is a weaker fit for teams that need custom architectures, cross-platform models, or cloud-scale training pipelines. Those teams will likely train elsewhere and convert to Core ML instead.
The honest bottom line: Create ML lowers the barrier to entry dramatically, but it is a training tool, not an app builder. Pair it with Xcode and the Apple Developer Program, keep your expectations scoped to its templates, and it earns its place in the toolbox.
Create ML comes with Apple's developer tooling for training on your own Mac, so there is no separate fee just to train a model. However, distributing an app that uses your model on the App Store requires the paid Apple Developer Program. Confirm current terms on Apple's official pages.
Yes. Create ML only trains the model and outputs a Core ML file. You build, test, sign, and submit the actual iOS app in Xcode. Create ML does not build or ship apps.
No. Create ML is macOS-only, both the app and the framework. There is no official Windows or Linux version and no first-party cloud training service.
Create ML trains models; Core ML runs them. You use Create ML to produce a model file, then use Core ML at runtime inside your app to make predictions on-device.
For its supported task types — image, text, sound, tabular, and similar — it can absolutely power production features. For custom architectures or cross-platform needs, teams often train elsewhere and convert to Core ML.