iOS On-Device AI Models
Production-ready guide for implementing on-device AI models in iOS apps using Apple's Foundation Models framework and MLX Swift.
When to Use This Skill
- Implementing local LLM inference in iOS apps
- Building chat interfaces with Foundation Models
- Integrating Vision Language Models (VLMs)
- Adding text embeddings or image generation
- Implementing tool/function calling with LLMs
- Managing multi-turn conversations
- Optimizing memory usage for on-device models
- Supporting internationalization in AI features
Core Principles
- Compatibility Gate - Foundation Models requires iOS 26, iPadOS 26, or macOS 26 or later, and an Apple Intelligence-capable device with Apple Intelligence enabled. Compare this floor with the deployment target and with every device the feature must support before you select a framework. Always check
SystemLanguageModel.default.availability at runtime before you create a session. Show a fallback UI for each unavailable state: .modelNotReady (model not ready, for example a download in progress), .appleIntelligenceNotEnabled (disabled in Settings), and .deviceNotEligible (hardware cannot run the model). A runtime check reports the state of one device. It cannot make the feature available on an OS version or device below the floor.
- Single-Flight Streaming - Give one isolation boundary ownership of each conversation's session, transcript, and generation task. Do not call the session while
isResponding is true; reject or queue the new send so that GenerationError.concurrentRequests never occurs in normal use. Stop and teardown cancel the owned task and await it before a new send starts. Check Task.checkCancellation() while you consume the stream. Commit a transcript turn only when the response completes.
- Session Persistence - Reuse LanguageModelSession across completed turns and keep partial streaming text separate from committed history.
- Memory Awareness - Use quantized models and monitor memory usage. iOS limits one app to a fraction of the device's total RAM. Reject any model whose weight files are about the size of, or larger than, the total RAM of the lowest-memory required device: it cannot load. Pick the smallest model that meets quality on that device and measure peak memory there. Give one
@Observable owner the load state, the loaded model, and generation. For a multi-gigabyte download, ask for consent, default to Wi-Fi, make the download resumable, and keep the feature in an explicit "not downloaded" state until the files are complete.
- Async Everything - Load models asynchronously, never block the main thread.
- Device Proof - Before calling the design viable, exercise support boundaries and generation lifecycle in focused tests, then verify a Release build on the oldest or lowest-memory required physical device, including offline operation when the product promises on-device behavior.
- Locale Support - Call
supportsLocale(_:) for each user locale before you create a session. Treat an unsupported locale as an unavailable state; do not use a prompt instruction such as "answer in " to work around it. For a supported locale, put the locale in the session instructions, not in each prompt. Test each locale branch on a physical device set to that locale.
- Typed Tools and Outputs - For structured output, define a
@Generable type and call respond(to:generating:); do not parse free text with regular expressions. For an app action the model needs, implement the Foundation Models Tool protocol with @Generable arguments and pass the tool when you create the session. The session invokes the tool. The view never calls the service on the model's behalf. A tool error surfaces from the session call as LanguageModelSession.ToolCallError. Catch it and show the user a normal failure state; never crash or show an empty result.
Quick Reference
Framework Comparison
Foundation Models (Apple's Framework)
MLX Swift (Advanced Features)
Shared (Both Frameworks)
Quick Decision Trees
Which framework should I use?
Is Foundation Models available on every OS version and device the feature
must support (iOS 26 or later, Apple Intelligence-capable hardware)?
├── No → Can a suitable MLX model meet the same device floor?
│ ├── Yes → MLX Swift (prove memory, latency, and output on that floor)
│ └── No → The requirements are infeasible; change the support contract
└── Yes → Do you need VLMs, image generation, or custom models?
├── Yes → MLX Swift (references/mlx-swift/)
└── No → Foundation Models (references/foundation-models/)
Where should I start?
New to on-device AI?
└── Start with Foundation Models:
1. Read framework-selection.md
2. Follow foundation-models/setup.md
3. Implement foundation-models/chat-patterns.md
Need advanced features?
└── Use MLX Swift:
1. Read framework-selection.md
2. Follow mlx-swift/setup.md
3. Choose pattern:
- Chat: mlx-swift/chat-patterns.md
- Vision: mlx-swift/vision-patterns.md
- Advanced: mlx-swift/advanced-patterns.md
Where should my model loading code live?
Is this model shared across features?
├── Yes → Create @Observable service in app/services/
└── No → Is it feature-specific?
├── Yes → Create @Observable class in feature/
└── No → Load inline with @State (simple cases only)
How should I handle conversations?
Foundation Models:
└── Reuse LanguageModelSession for context
(references/foundation-models/chat-patterns.md #multi-turn)
MLX Swift:
└── Implement custom context management
(references/mlx-swift/chat-patterns.md)
What generation parameters should I use?
What's the use case?
Factual answers (summaries, facts)
└── temperature: 0.1-0.3
Balanced (chat, Q&A)
└── temperature: 0.6-0.8
Creative (storytelling, ideas)
└── temperature: 0.9-1.2
See references/shared/best-practices.md for details
Resources
