Core ML Deployment

How to run a model you've converted to Core ML inside your app: getting it onto the device, loading it fast, running predictions safely under Swift 6, and using the newer model features such as a persistent KV cache for language models and several adapters in one model file.

When to Use

Use this skill when you're:

• Loading a converted model and running predictions in an app
• Fixing a first prediction that's much slower than the rest
• Running predictions from several tasks or a background queue under Swift 6
• Choosing compute units (CPU, GPU, Neural Engine)
• Building a language model's token loop with a KV cache that persists between predictions
• Shipping several adapters or tasks in one model asset
• Chaining models, or decoding a model's output, with MLTensor

Example Prompts

Questions you can ask your assistant that will draw from this skill:

• "How do I load my Core ML model so the first prediction isn't slow?"
• "Swift 6 says sending my MLModel risks a data race. How should I structure predictions?"
• "How do I run predictions on many images in parallel without Swift 6 data-race errors?"
• "How do I keep a KV cache between predictions for my language model?"
• "Can I ship two LoRA adapters in one Core ML model?"
• "Should I set computeUnits to the Neural Engine only?"

What This Skill Provides

• Model assets – Bundling a compiled model, compiling a downloaded one on the device, and where to keep the result
• Load, cache, and pre-warm – Why the first load is slow, and how to pay that cost before the user needs the model
• Predictions under Swift 6 – Why MLModel can't be shared across tasks, and three structures that compile cleanly: an actor that owns the model, one model instance per parallel lane, and a job that owns its model for one batch of async predictions
• Compute units – Choosing among CPU, GPU, and Neural Engine, and the iOS 27 entitlement for running inference in the background
• Stateful models – Keeping a transformer's KV cache in model state across predictions (iOS 18 and later)
• Multifunction models – One asset holding several functions, such as adapters that share base weights (iOS 18 and later)
• MLTensor – Pre- and post-processing between models without hand-written array loops (iOS 18 and later)
• Anti-patterns and a checklist – Including the concurrency shortcuts Swift 6 rejects

Documentation Scope

This page documents the coreml-deployment skill (in the axiom-ai suite), which your assistant loads when you run a Core ML model in an app. Language-model-scale models on the iOS 27 runtime are covered by Core AI.

Related

• ios-ml – The hub that routes Core ML work to each stage
• coreml-conversion – Producing the model you deploy, including stateful models
• coreml-compression – Shrinking a model that's too large or slow to deploy
• coreml-ref – API signatures, availability, and which types are Sendable
• coreml-diag – When a model fails to load, runs slowly, or grows memory

Resources

WWDC: 2023-10049, 2024-10161, 2024-10159

Docs: /coreml/mlmodel, /coreml/mlmodelconfiguration, /coreml/mlstate, /coreml/mltensor, /coreml/mlmodelasset