Core ML

On-device machine-learning inference on Apple platforms — image classification/detection, NLP, audio, and custom models, plus on-device fine-tuning. The deep API reference — model integration, the MLModel API, MLFeatureProvider, async predictions, Vision integration, configuration/optimization, stateful models, encryption, and version compatibility — lives in references/guide.md. This file is the decision and discipline layer: read it first, open the guide for specifics.

Dials

Set these explicitly at the start; they change what "correct" means.

1. TASK — image (classification/detection/segmentation; drive it through Vision, not raw MLModel) · text (NLP via NLModel/Natural Language) · tabular-or-custom (raw MLModel + MLFeatureProvider). If the task is text generation, stop and use foundation-models instead.
2. COMPUTE_UNITS — all (default; OS picks CPU/GPU/Neural Engine) · cpuAndNeuralEngine (NN-heavy models) · cpuAndGPU · cpuOnly (background/deterministic timing). Set on MLModelConfiguration at load, never per-prediction.
3. LOADING — xcode-bundled (default; drop in .mlpackage/.mlmodel, Xcode compiles + generates the class) · runtime-download (download .mlmodel, MLModel.compileModel(at:), then MLModel.load(contentsOf:) — for large or swappable models).

When to use

Building or reviewing any on-device inference path: loading a model, configuring compute units, running predictions (sync or async), wiring a model into Vision for images, encrypting a bundled model, running a stateful generator, or on-device personalization (MLUpdateTask). If the goal is generating or reasoning over text, that's foundation-models, not Core ML.

Core rules