Tools for Swift Devs to Access Apple Docs with LLMs

When to use Claude versus the on-device model? Apple Foundation Models Apple's on-device model is fast, private, and works offline, but it is sized for lightweight tasks. Escalate to Claude when you need larger context, frontier reasoning, or server-side tools such as web search and code execution. Because both use the same LanguageModelSession API, you can switch by swapping the model: argument. https://lnkd.in/eywWzsGW

WWDC26 Developer Breakdown — Part 3 of 7 Apple opened its AI stack to developers in a big way. Two frameworks to understand — they solve different problems. 1. Foundation Models Framework A native Swift API for language models in your app. The headline: one LanguageModelSession API, swap providers without rewriting architecture. Supported providers: → Apple Foundation Models (on-device) → Private Cloud Compute (server-side Apple models) → Third-party cloud models — Claude, Gemini, others via Swift packages → Core AI (your custom on-device models) → MLX (Hugging Face community models) New capabilities: → Multimodal input — images + text in the same prompt → On-device Vision tools (OCR, barcode) callable by the model → Dynamic Profiles — swap models, tools, and instructions mid-session → Evaluations framework — test AI behavior beyond what unit tests catch Small Business win: Apps with fewer than 2M first-time downloads get free Apple Foundation Models on Private Cloud Compute. 2. Core AI Framework Purpose-built for bringing YOUR models on-device. → Optimized for Apple Silicon (CPU, GPU, Neural Engine) → Zero server dependency, zero token costs → Ahead-of-time model compilation in Xcode → Core AI Instruments for profiling → Plugs into Foundation Models via CoreAILanguageModel When to use which? Foundation Models → system LLM, Siri-class intelligence, Claude/Gemini via Swift API Core AI → custom vision models, fine-tuned domain models, bundled generative AI The unified API means your app architecture stays stable even as model providers evolve. Part 4 next: App Intents — making your app a first-class citizen in Siri AI. Building any AI features in your iOS app this year? #SwiftLang #MachineLearning #AppleIntelligence #iOS27 #SoftwareEngineering

📱 I used to think Apple would end up with one AI model living on your phone. After WWDC, I think I was wrong - and what's actually happening is better. WWDC 2026 was Monday. One theme ran through the entire keynote: AI is moving onto the device. 🧠 Why on-device wins • Cloud LLM tokens get expensive fast. At scale, every call is a line item on your P&L. • An on-device model is free per call, private, and works offline. • For an indie app, that doesn't just trim costs - it flips the unit economics. ⚙️ And it's finally buildable Apple's Foundation Models framework now gives you: • A direct API to the 3B on-device model • Multimodal input now (images + text), and it even runs on Linux • A Python SDK and an fm CLI you can pipe shell scripts through • A new LanguageModel protocol where Apple's model, Gemini and Claude all implement one interface - swap providers in a single line 🔥 The part that changed my mind Apple didn't ship one model. It shipped on-device AI at every layer in one keynote: • The OS - Siri rebuilt around it • The apps framework - Foundation Models • The IDE itself - Xcode 27 with on-device code completion, no cloud round-trip So the future isn't "one model on the phone." It's the model layer becoming a swappable, on-device-first interface. Small model on the device, the cloud only for the heavy lifts, third-party models dropping in through one protocol. These models are Swift-native, so us React Native devs will need a bridge to call them from JS. That's exactly the work I'm in right now - I'm building on-device AI into EndedUp, one of my own apps. This one isn't commentary for me, it's my current sprint. Are you betting on-device, or staying in the cloud? 👇 #ondeviceai #wwdc2026 #appleintelligence #reactnative #mobiledev #iosdev #ai #foundationmodels #indiedev #javascript

I like how Apple is moving toward MoE with only a few billion active parameters, and I think their query-level routing approach is an interesting way to deal with the low NAND-DRAM bandwidth bottleneck for on-device MoE. But I’m not convinced by the Evaluation section: 1. No comparison with 3rd-party open models. For an on-device model, I’d expect comparisons against models like Gemma, Qwen, Granite, and LFM, especially when the reported results are all from Apple’s internal benchmarks. Only comparing against Apple’s own previous generation makes it hard to understand where AFM 3 actually stands. 2. No agentic or tool-use benchmarks. Apple mentions agentic tool use and complex reasoning, but the eval section does not report standard agent/tool-use benchmarks. Besides internal evals, why not include BFCL, τ-bench, or similar? For mobile, coding and math may not be the most representative tasks, but agentic tool-use should be one of the most important things to measure. 3. AFM 3 Core Advanced is barely evaluated as a general model. The 20B sparse model is arguably the most interesting part of the release, but its public evals focus on feature-level TTS / dictation rather than general text, reasoning, multilingual, or tool-use benchmarks. Article: https://lnkd.in/gsqXZ-7i

The bug didn't change. The AI didn't change. I turned one dial — and that cracked it. Our iOS app had a bug: AI-assisted calls failed on Mac. Connect, then die 2-3 seconds later. Regular VoIP calls were fine. I put Claude Code on it. At its normal setting, it saw the Xcode console flooded with scary Apple audio errors and concluded the Mac's audio engine was broken. It wrote fixes, then recommended we rewrite our audio layer or drop MacOS support. It even drafted a GitHub issue. All confidently, thoroughly wrong. Then I switched it to max effort (Claude Code's /effort, running Opus) — same model, same context, nothing added. That alone reframed everything. It finally used a fact that had been in front of it the whole time: regular calls work fine on the same Mac, same audio stack. So it checked whether those scary errors also appeared in the working case. They did — harmless noise on every successful call. A red herring. The real cause: AI calls routed through CallKit; regular ones didn't. On Mac (no push notifications) CallKit was auto-ending the call. The fix was a few lines. Worked first try. The lesson isn't "AI is smart now." It's that reasoning depth is a dial, not a constant — and at low depth it fails like a rushed engineer: latch onto the loudest signal, build a theory, defend it. Two takeaways: 1. Before trusting an error, check if it also shows up when things work. 2. When your tools let you spend more reasoning, spend it. The cheap answer and the correct answer often look equally confident. #ClaudeCode #Claude #AIcodingTools #debugging  #iOSdev