Machine Learning

Skills for deploying and running machine learning models on Apple devices using CoreML and related frameworks.

Available Skills

iOS ML

The hub for on-device machine learning and speech-to-text. It routes your question to the right skill for each stage of a Core ML model's life, to Core AI for language-model-scale models, or to speech.

Core ML Conversion

Convert a trained PyTorch, TensorFlow, or Keras model with coremltools, choose a deployment target, convert stateful models, and check that the converted model matches the original.

Core ML Compression

Shrink a model with quantization, palettization, or pruning, decide between post-training and training-time compression, and re-measure accuracy at each step.

Core ML Training

Train a new model with Create ML, or personalize a deployed model on the user's device with MLUpdateTask, and know its format limitation before you build.

Core ML Deployment

Run a converted model in your app: loading and pre-warming, compute units, Swift 6-safe prediction, stateful KV-cache models, multifunction adapters, and MLTensor pipelines.

ML Training Paths

Pick the right training toolchain before you build — disambiguates Foundation Models adapters, Create ML, MLUpdateTask personalization, coremltools conversion, and MLX, and flags the format and toolchain traps that make a trained artifact fail to load.

Core AI

Core AI (iOS 27) — the on-device inference framework behind Apple Intelligence, now open to apps. Convert PyTorch models to .aimodel, run them from Swift across CPU, GPU, and Neural Engine, and back a LanguageModelSession with your own model.

Speech

On-device speech-to-text with SpeechAnalyzer (iOS 26 and later): file and live transcription, model assets, the limit on simultaneous analyses, and the iOS 27 audio input providers.

Available References

• Core ML Reference – MLModel, configuration, compute devices, prediction and Sendability, MLState, MLTensor, MLComputePlan, and the coremltools index

Available Diagnostics

• Core ML Diagnostics – Load errors, slow first load or predictions, memory growth, device and Simulator differences, and accuracy loss after compression

Example Prompts

• "How do I convert a PyTorch model to CoreML?"
• "My CoreML model is too large, how do I compress it?"
• "How do I implement speech-to-text with SpeechAnalyzer?"
• "Model inference is slow, how do I optimize it?"