Train custom ML on your laptop, deploy it to an iPhone

What it does

Turi Create is a Python framework that wraps common machine learning tasks—image classification, object detection, recommender systems, style transfer, and more—into high-level, task-oriented APIs. You bring the data, it picks the algorithm, trains the model, and exports a Core ML file ready for iOS, macOS, watchOS, or tvOS. It is designed to run on a single machine and claims to handle large datasets without requiring a distributed cluster.

The interesting bit

The entire workflow is optimized for the Apple stack: train on macOS or Linux, but every example and highlighted feature funnels the result into Core ML. It is Apple’s attempt to own the full path from training to on-device inference.

Key highlights

• Task-focused API: choose a job like “image classifier” or “recommender” instead of tuning raw algorithms.
• Direct export to Core ML for on-device inference across Apple platforms.
• Handles text, images, audio, video, and sensor data in the same toolkit.
• Optional GPU acceleration for specific tasks, with the README claiming 9–13× speedups on supported hardware.
• Runs offline on a single machine; minimum requirement is 4 GB RAM and an x86_64 chip.

Caveats

• Windows support is WSL-only, not native.
• Python support stops at 3.8, and 2.7 is still listed, which feels dated.
• No ARM builds mentioned; x86_64 is explicitly required.

Verdict

A practical choice if you are building Apple-native apps and want on-device ML without wrestling with TensorFlow or PyTorch directly. Look elsewhere if you are targeting Android, servers, or non-x86 hardware.