WWDC 2026 just changed how every iOS developer builds AI features. Apple's new LanguageModel protocol in the Foundation Models framework lets you swap between Apple's on-device Neural Engine, Apple's Private Cloud Compute, Google Gemini, and Anthropic Claude — by changing ONE Swift Package dependency. No code changes in your session logic. Here's what this means in practice: → Prototype with Apple's on-device model (free, private, no network call) → Escalate to Private Cloud Compute (Apple Silicon servers, stateless, auditable by security researchers) → Route the heaviest queries to Google Cloud's Blackwell B200 GPUs running a custom 1.2-trillion-parameter Gemini model → Or swap to Anthropic's Claude via their Swift package — same protocol, same API surface Apple's privacy layer proxies every handoff: queries are anonymized, stripped of Apple ID linkage, and tokenized before reaching Google's infrastructure. Federighi said in the live Q&A: "None of the models Google deploys to its customers. Completely private to you. Never stored." The new built-in tools are equally significant: → BarcodeReaderTool OCRTool backed by Vision framework → Spotlight-powered local RAG — no vector database infrastructure needed → 32K context window on the server model, configurable reasoning levels The catch: if your app hasn't implemented App Intents (available since iOS 16), the new Gemini-powered Siri won't see it. SiriKit is formally deprecated. Migration window is 2-3 years, but apps that don't migrate won't surface in the rebuilt Siri experience this fall. This is the most architecturally significant iOS cycle in years. Apple didn't build a model — they built a router with privacy guarantees that lets developers use any model. techtimes.com/articles/31803…

Claude code has been incredibly useful for me to rapidly try out Core AI, and I was able to quickly implement a LanguageModel protocol for Open Responses compatible endpoints. So now I have always on models, Core AI embedded model, and cloud models! github.com/RichNasz/SwiftOpe…

Great article! Well done @malekoo 🙏 I loved reading it! Playing with Swift and LanguageModel protocol too here!

This week in AI Dev: Apple's LanguageModel protocol, Claude Fable 5 lab-only Mythos 5, OpenAI Migrate to Codex, Google Antigravity, Cursor Composer 2.5, Niteshift, Cohere North Mini Code, DiffusionGemma, Stack Overflow for Agents, WebMCP, Mastercard AP4M, and xAI's SPCX. Catch up on all the news everydev.ai/p/news-weekly-ai…

Apple just made every iOS AI feature a Swift one-liner — and it changes the entire mobile AI landscape. WWDC 2026 shipped the Foundation Models framework as a single native Swift API that can route to Apple's on-device Neural Engine model, Private Cloud Compute, Anthropic Claude, or Google Gemini — all from the same call site. No API keys. No provider-locked code. Here's what actually landed: 1. LanguageModel protocol — any provider (Claude, Gemini, open-source) implements a Swift package. Your app swaps providers by changing a dependency, not rewriting logic. 2. Image input on-device — pass images alongside text. The model can call Vision framework tools (OCR, barcode reading) directly. No cloud vision API needed. 3. Dynamic Profiles — swap tools and instructions at runtime. This is Apple's foundation for multi-agent workflows inside a single app session. 4. Private Cloud Compute is FREE for Small Business Program devs (<2M downloads). No cloud API cost for Apple Foundation Models. 5. Core AI framework — run full-scale LLMs locally on Apple Silicon. Optimized for unified memory and Neural Engine. You deploy your own custom model on device. 6. Xcode 27 agentic coding — Claude, Gemini, and OpenAI agents built into the IDE. Agents can write tests, run Playgrounds, interact with the Simulator, and validate their own work autonomously. The architectural shift nobody's talking about: Apple open-sourcing this framework later this summer. That's unprecedented for Cupertino. They're betting the mobile AI stack will standardize around Swift, not Python. My take: the LanguageModel protocol is the App Store moment for mobile AI. Any developer shipping on iOS 27 can now build AI features without managing infrastructure. The barrier just collapsed. developer.apple.com/newsroom…

Started this week with Xcode 27 and so far so good. Posting a LanguageModel protocol adapter for Open Responses endpoints very soon. Testing on my Vision Pro connecting to MacBook serving models using lm studio without issues so far.

Ho finalmente completato il corso di AI Agentic Applications Masterclass di Data Masters! Partivo più o meno da zero ed ho dovuto faticare un pò. Ma come si dice: no pain no gain! #AI #languagemodel #generativeai #langchain #rag #agenticapplications data-masters-srl.verified.cv…

macOS and iOS apps with Core AI with various LanguageModel including MLX clearly 😂 But just as learning experiment for now.

AIニュースピックアップ ｜ 2026/6/12 --- GoogleがDiffusionGemma-26BをApache 2.0でオープンソース公開した。 自己回帰ではなくトークンを並列でデノイズする「テキスト拡散」アーキテクチャで、H100上で毎秒1,000トークン超、RTX 5090でも700トークン超を出す。 量子化後は18GB VRAMに収まり、ハイエンドのコンシューマGPUで動く。品質は標準Gemma 4より落ちるとGoogle自身が明記しているが、速度を優先したい用途向けにvLLM・Transformers・MLXで動かせる。 --- OpenAIがVisaと提携し、AIエージェントがユーザーに代わって購入できる仕組みを発表した。 ChatGPTのエージェントが、ユーザーの設定した支出上限・承認閾値・加盟店制限の範囲内でVisaカードで決済できるようになる。 Visaのトークン化認証とリアルタイム不正検知を組み合わせた設計。今のところ発表のみでリリース時期・価格は未定。 --- WWDC 2026でAppleのFoundation ModelsフレームワークにLanguageModelプロトコルが追加された。 Swift Package Managerの依存関係を更新するだけで、アプリのコードを変えずにオンデバイスモデル・Claude・Geminiを切り替えられる。 AnthropicとGoogleがすでにSwiftパッケージを公開しており、iOS 27・macOS 27・iPadOS 27・visionOS 27に対応。iOSアプリへのLLM組み込みで、プロバイダーを後から差し替えられる構造になる。 --- GoogleがGemini 3.5 Proを6月中にリリースすると予告している。 コンテキストウィンドウ200万トークン、推論モード「Deep Think」搭載。Vertex AIでは限定プレビューが始まっており、価格は入力$2〜3・出力$12〜18/百万トークン見込み。 Claude Fable 5やGPT-5.5との直接比較が今月中に始まる。

Follow-up on what's doing the magic: the framework, not the models. Two protocols. LanguageModel declares capabilities. LanguageModelExecutor does the work: prewarm, translating the Transcript to each engine's native format, streaming. AFM and Qwen via MLX plug into the same contract. The session owns the state. Append-only, coupled to the KV cache, and in macOS 27 the transcript is mutable with clear rules: append and the cache survives, rewrite history and you pay a fresh prefill. Dynamic Profiles sits on top. That's why per-prompt AFM vs MLX routing works and the TTFT numbers hold mid-conversation. Apple shipped an on-device inference orchestrator and called it a framework update. #WWDC26 developer.apple.com/videos/p…

#𝐆𝐈𝐁𝐑𝐘 𝐁𝐫𝐢𝐧𝐠𝐬 𝐂𝐥𝐚𝐮𝐝𝐞 𝐅𝐚𝐛𝐥𝐞 𝟓 𝐭𝐨 𝐌𝐀𝐍𝐀 𝐒𝐭𝐮𝐝𝐢𝐨 𝙍𝙚𝙖𝙙 𝙈𝙤𝙧𝙚: itbusinesstoday.com/news-art… #AIagents #Customersupport #existingworkflow #generativeAI #GIBRY #itbusinesstoday #LanguageModel #Marketingteams #News

Apple ha presentado en la WWDC 2026 la tercera generación de sus Apple Foundation Models, AFM 3. La familia incluye cinco modelos: AFM 3 Core, AFM 3 Core Advanced, AFM 3 Cloud, ADM 3 Cloud para imagen y AFM 3 Cloud Pro. AFM 3 Core es el modelo local base, de unos 3.000 millones de parámetros. AFM 3 Core Advanced es un modelo multimodal de 20.000 millones de parámetros diseñado para ejecutarse en dispositivos Apple Silicon mediante una arquitectura sparse. En AFM 3 Core Advanced, el modelo completo puede residir en memoria flash, mientras que solo una parte de los expertos se carga en DRAM para cada petición. Apple denomina a esta técnica Instruction-Following Pruning, IFP. La selección de expertos se realiza a partir de la instrucción del usuario y puede actualizarse durante la generación. En la parte cloud, Apple introduce AFM 3 Cloud y AFM 3 Cloud Pro dentro de Private Cloud Compute. AFM 3 Cloud Pro está orientado a tareas de razonamiento más complejas y a escenarios con uso de herramientas. Apple también indica que este modelo se ejecuta sobre GPUs NVIDIA en Google Cloud, dentro de su infraestructura de computación privada. El Foundation Models framework permitirá a los desarrolladores usar modelos locales de Apple, modelos de Private Cloud Compute y modelos de terceros que implementen el protocolo LanguageModel. Apple publicará próximamente un informe técnico completo, con evaluaciones y benchmarks actualizados. machinelearning.apple.com/re…

Apple Foundation Models の LanguageModel プロトコルを試してます。SPM の依存を1行変えるだけでオンデバイス → Claude → Gemini を切り替えられる設計で、session コードは変えない。guided generation も tool calling も同一インターフェースで動く。逆に Swift ファーストなのでサーバーサイド Python から呼ぶには一工夫いる。iOS アプリへの LLM 組み込みのモデルルーティング設計がかなりシンプルになりそうです。

So, when is @OpenAIDevs releasing the OpenAI LanguageModel SPM?

I like the LanguageModel protocol added to the Foundation Models Framework, because you can plug in any inference engine to it even your custom one! I really want to give this a try, combined with the new Evaluation framework 🧐

The whole thing now hinges on a new 'LanguageModel' protocol. A 'LanguageModelSession' is what you talk to. It's backed by any type conforming to LanguageModel - that protocol is the abstraction that lets any model drive a session. Existing models like SystemLanguageModel and PrivateCloudComputeLanguageModel already conform to it. Pick the model, the session API stays identical.

STEELMAN cont - The LanguageModel protocol, the zero-cost PCC for small businesses, and the React Native integration are not model innovations — they are ecosystem plays that ensure developers build against Apple's API surface regardless of which model sits behind it, making the specific frontier model an interchangeable commodity Apple can swap at will - The hybrid architecture's privacy gaps are real, but they are also irrelevant to the strategic outcome: Apple's lock-in has never been about perfect privacy — it has been about making the cost of switching higher than the cost of staying, and WWDC 2026 just raised that switching cost to the level of an entire AI development pipeline

The Foundation Models framework commoditizes the LLM backend - The LanguageModel protocol is a public Swift interface that abstracts the inference provider from application logic — developers build against Apple's standard, not a specific model - A team can prototype with the free-to-execute AFM 3 Core, then update a single Swift Package Manager dependency to point to Anthropic's Claude or Google's Gemini (via the Firebase Apple SDK) without touching core application code

I just used GPT5.5 and created LanguageModel implementation for OpenAI models! Took a few extra prompts to get streaming and reasoning effort but it looks like it works!