Apple used its 2018 developer conference to make machine learning feel less like a specialist research discipline and more like a standard part of application development. Create ML gave developers a Swift-based route to training models, while Core ML 2 improved the framework used to run those models inside Apple applications. Together, the tools expressed a clear strategy: simplify model creation, shrink deployment costs and keep more intelligence on the device.
The announcement mattered because most developers did not have dedicated machine-learning infrastructure or large research teams. Apple wanted an application designer already working in Xcode to experiment with classification and prediction without adopting an entirely separate toolchain.
Training Models in a Familiar Environment
Create ML was designed to work with Swift and Xcode Playgrounds. Developers could provide labelled examples, select an appropriate task and train a model through interfaces that reduced the amount of specialised code required. The aim was not to replace advanced research frameworks, but to make common machine-learning jobs practical for a much wider audience.
Apple highlighted a language-learning application that had previously needed 24 hours to train an image model using 20,000 examples. With Create ML, the reported training time fell to 48 minutes on a MacBook Pro and 18 minutes on an iMac Pro. The resulting model also dropped from roughly 90MB to 3MB.
Those figures represented a carefully selected case, so they were not a guarantee for every dataset. Still, they illustrated what framework-level optimisation could mean for small teams. A shorter training cycle allows more experiments, while a compact model is easier to distribute with an app and less demanding on a phone’s storage.
Core ML 2 Focused on Efficient Inference
Core ML had arrived a year earlier as Apple’s framework for integrating trained models into applications. The second version concentrated on speed and size. Apple said batch prediction made it up to 30 percent faster for relevant workloads, while quantisation could reduce model size by as much as 75 percent.
Quantisation represents model values with lower numerical precision. When applied carefully, it reduces memory and computation while preserving acceptable accuracy. That trade-off is especially useful on mobile hardware, where power, thermal limits and storage matter as much as peak performance.
Core ML supported more than neural networks. Developers could deploy tree ensembles, support vector machines and generalised linear models alongside deep-learning architectures. Underneath, the framework used technologies such as Metal and Accelerate to direct work efficiently to the CPU and GPU.
Privacy Became a Technical Advantage
Apple repeatedly emphasised that Core ML models could run locally. Data did not have to leave the device for analysis, which reduced network delay and allowed features to work offline. It also supported Apple’s growing argument that privacy could be built into the architecture rather than added only through policy.
On-device processing is not automatically private. Applications still determine what they collect and share, and a poorly designed model can expose sensitive information. Yet local inference gives developers the option to keep photos, text or behavioural signals away from remote servers.
The approach also limited what a model could do. Cloud systems could draw on larger models and continuously updated information, while a phone had a finite memory and energy budget. Apple’s response was to make models smaller and execution more efficient instead of assuming every intelligent feature needed a data-centre connection.
A Bridge to the Wider Ecosystem
Create ML and Core ML did not isolate Apple developers from other machine-learning tools. Models built in frameworks such as TensorFlow could be converted into Core ML formats. That bridge allowed teams to train with familiar research software and deploy through Apple’s optimised runtime.
The combination created a two-level platform. Specialists could continue building sophisticated models elsewhere, while application developers gained a simpler route for common tasks. Both groups could target the same device ecosystem.
This was strategically important. A machine-learning framework becomes more valuable when developers use it to create features Apple would not build itself. Image recognition, text classification and recommendations could spread across thousands of applications, increasing the usefulness of the hardware.
Machine Learning as an Everyday Developer Tool
In 2018, Create ML and Core ML 2 were early steps rather than a complete no-code AI platform. Training quality still depended on representative data, careful evaluation and an understanding of model limitations. Simplifying an interface did not eliminate the responsibility to test bias, accuracy and failure cases.
What Apple changed was the starting point. Developers no longer had to treat machine learning as a remote laboratory exercise. They could train a compact model on a Mac, add it to an application and run it privately on a user’s device.
That workflow anticipated a broader shift toward edge AI. By tying accessible training tools to efficient local inference, Apple made machine learning part of ordinary application engineering—and turned its control of chips, operating systems and developer frameworks into a practical platform advantage.
