Swift AI for macOS (2026)

Building AI features into native macOS (and iOS) apps with Swift gives you two paths: call cloud LLM APIs from Swift, or run models on-device with Apple's frameworks for privacy and offline use. This guide covers both and when to choose each.

Two approaches

Cloud APIs from Swift: call OpenAI/Anthropic over HTTPS with URLSession. Most capable models, simplest to start, but needs network and sends data off-device.

On-device: run models locally via Apple's frameworks (Core ML, the Foundation Models / Apple Intelligence APIs) or local runtimes (llama.cpp via MLX/Metal). Private, offline, no per-call cost — but limited to smaller models.

Calling a cloud LLM from Swift

swift
struct ChatReq: Encodable { let model: String; let messages: [[String:String]] }
func ask(_ prompt: String) async throws -> String {
    var req = URLRequest(url: URL(string: "https://api.openai.com/v1/chat/completions")!)
    req.httpMethod = "POST"
    req.setValue("Bearer \(apiKey)", forHTTPHeaderField: "Authorization")
    req.setValue("application/json", forHTTPHeaderField: "Content-Type")
    req.httpBody = try JSONEncoder().encode(
        ChatReq(model: "gpt-4o", messages: [["role":"user","content":prompt]]))
    let (data, _) = try await URLSession.shared.data(for: req)
    // decode choices[0].message.content ...
    return String(decoding: data, as: UTF8.self)
}

Use streaming (URLSession bytes / SSE) for responsive UIs — same idea as Streaming AI with SSE.

On-device with Apple frameworks

Apple Intelligence's Foundation Models framework exposes an on-device LLM to Swift apps for summarization, generation, and tool calling — private and offline, ideal for system-integrated features. Core ML runs your own converted models; MLX/llama.cpp run open models on Apple Silicon (Metal). For picking a small local model, see local LLM comparison.

Choosing

Need the most capable model / complex reasoning? Cloud API.

Privacy, offline, no per-call cost, system integration? On-device.

Hybrid: on-device for quick/private tasks, cloud for heavy lifting — a common pattern.

FAQ

Can I run an LLM fully on a Mac? Yes — via Apple's on-device models, Core ML, or MLX/llama.cpp on Apple Silicon. How do I call OpenAI from Swift? URLSession POST to the chat endpoint; stream for responsive UIs. Why on-device? Privacy, offline use, and no per-request cost — at the price of smaller models. Best of both? Hybrid: on-device for light/private tasks, cloud for the hard ones.

Summary

For AI in Swift/macOS, call cloud LLMs with URLSession for maximum capability, or use Apple's on-device frameworks (Foundation Models, Core ML, MLX) for private, offline, zero-cost inference. Many apps blend both — local for quick private tasks, cloud for heavy reasoning.

---

*Last updated: June 2026. Verify against Apple's developer documentation.*