iOS Machine Learning

Navigation hub for on-device machine learning and speech-to-text work. Routes to the right specialized skill based on whether you're converting models, optimizing inference, debugging, or transcribing audio.

When to Use

Use this skill when you're:

• Converting a PyTorch or TensorFlow model to CoreML
• Deploying a custom ML model on iOS, iPadOS, macOS, watchOS, or tvOS
• Compressing a model with quantization, palettization, or pruning
• Implementing KV-cache for a custom large language model
• Stitching multiple models together with MLTensor
• Adding live or file-based speech transcription
• Debugging slow inference, load failures, or accuracy drops after compression

Boundary with Foundation Models

ML and AI sound interchangeable but route to different skills:

What you're doing	Skill
Running your own model on-device	This skill (CoreML)
Using Apple's built-in on-device LLM	Foundation Models
Specializing the built-in LLM with a custom adapter	Foundation Models Adapters
Image analysis, pose detection, text recognition	axiom-vision skills

Rule of thumb — if you're converting, compressing, or deploying your own model, you want CoreML. If you're calling Apple's LanguageModelSession, you want Foundation Models.

Example Prompts

Questions you can ask Claude that will draw from this skill:

• "How do I convert my PyTorch model to CoreML?"
• "My model is 5 GB. How do I compress it to fit on iPhone?"
• "Should I use CoreML or Foundation Models for text generation?"
• "How do I implement KV-cache for my LLM?"
• "Why is my CoreML inference slow on the first call?"
• "What's MLTensor and when should I use it?"
• "Add live transcription with SpeechAnalyzer to my app."
• "My compressed model has bad accuracy. What went wrong?"

What This Skill Provides

This skill routes to four specialized resources depending on your task:

• CoreML implementation patterns — model conversion, compression strategies, stateful models with KV-cache, multi-function models (LoRA adapters), MLTensor pipeline stitching, async concurrent prediction, compute unit selection
• CoreML API reference — CoreML Tools Python API, MLModel lifecycle, MLTensor operations, MLComputeDevice availability, state management APIs, performance reports
• CoreML diagnostics — load failures, slow inference, memory pressure, accuracy degradation after compression, compute unit problems
• Speech transcription — SpeechAnalyzer setup (iOS 26+), SpeechTranscriber configuration, live vs file transcription, volatile vs finalized results, AssetInventory model management

Related

• coreml — implementation patterns for model conversion, compression, and deployment
• coreml-ref — comprehensive CoreML API reference with method signatures
• coreml-diag — troubleshooting when models fail to load, predict slowly, or lose accuracy after compression
• speech — SpeechAnalyzer and SpeechTranscriber patterns for live and file transcription
• foundation-models — use Apple's built-in on-device LLM instead of a custom model when the task fits
• foundation-models-adapters — train a custom adapter on Apple's base model instead of deploying a separate CoreML LLM

Resources

WWDC: 2023-10047, 2023-10049, 2024-10159, 2024-10161, 2025-256

Docs: /coreml, /coreml/mlmodel, /coreml/mltensor, /speech, /speech/speechanalyzer