iOS Machine Learning
The hub for running your own machine learning models on Apple devices with Core ML, and for on-device speech-to-text. It sends each question to a dedicated skill for conversion, compression, training, deployment, diagnostics, or speech.
When to Use
Use this skill when you're:
- Converting a PyTorch or TensorFlow model to Core ML
- Deploying a custom ML model on iOS, iPadOS, macOS, watchOS, or tvOS
- Compressing a model with quantization, palettization, or pruning
- Implementing a 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 they lead to different skills:
| What you're doing | Skill |
|---|---|
| Running your own model on-device | This skill (Core ML) |
| Using Apple's built-in on-device LLM | foundation-models |
| Specializing the built-in LLM with a custom adapter | foundation-models-adapters |
| Bringing an LLM-scale or transformer model on-device (iOS 27) | core-ai |
| Image analysis, pose detection, text recognition | vision |
Rule of thumb If you're converting, compressing, or deploying your own model, you want Core ML. If you're calling Apple's LanguageModelSession, you want Foundation Models. If you're bringing an LLM-scale or transformer model, use Core AI (iOS 27).
Example Prompts
Questions you can ask your assistant 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?"
- "How do I use Core AI to serve my own on-device model to Foundation Models?" (iOS 27)
What This Skill Provides
- Routing – Sends each question to the skill for its stage: conversion, compression, training, deployment, API lookup, diagnostics, or speech
- The boundaries – When a task belongs to Core ML, Core AI (iOS 27), Foundation Models, or Vision
- Core AI pointer – When a language-model-scale model belongs in Core AI instead of Core ML
Documentation Scope
This page documents the ios-ml skill (in the axiom-ai suite), the hub your assistant starts from for on-device machine learning and speech-to-text. The work itself lives in the skills listed below.
Related
- coreml-conversion – Converting a trained model with
coremltools - coreml-compression – Shrinking a model with quantization, palettization, or pruning
- coreml-training – Create ML training and on-device personalization
- coreml-deployment – Running a model in your app under Swift 6
- coreml-ref – Core ML API reference
- coreml-diag – Troubleshooting load failures, slowness, memory, and accuracy loss
- speech –
SpeechAnalyzerandSpeechTranscriberfor live and file transcription - training-paths – Choosing among Create ML,
MLUpdateTask, adapters, and the other training toolchains before you build
Resources
WWDC: 2023-10047, 2023-10049, 2024-10159, 2024-10161, 2025-277
Docs: /coreml, /coreml/mlmodel, /coreml/mltensor, /speech, /speech/speechanalyzer