AI Agents

Apple Opens Its Foundation Models Framework to Gemini, Claude, and Beyond

Published Jun 11, 2026AI Agents

At WWDC 2026, Apple turned its on-device Foundation Models framework into a vendor-neutral, agentic developer platform — letting any provider plug into one Swift API, giving small developers free cloud inference, and promising to open-source the framework this summer.

The story

At WWDC 2026 on June 8, Apple did something more consequential than the Siri headlines suggested: it turned its on-device-first Foundation Models framework into a vendor-neutral, agentic developer platform. The centerpiece is a new public LanguageModel protocol that lets any model provider — Google's Gemini and Anthropic's Claude were named explicitly — plug into the same Swift API that already drives Apple's own on-device models [6][7].

For practitioners, the pivot that matters is the protocol-and-framework change, not the marketing. Apple is making model routing, fallback, and agentic scaffolding an operating-system concern rather than custom glue code [2][3].

What was actually announced

The headline change is portability. The on-device Apple model (SystemLanguageModel) and cloud models now sit behind the same Swift API surface. According to Firebase, switching from the on-device model to Gemini is "a small code change: swap the model instance" — initialize a LanguageModelSession with a GeminiLanguageModel instead of a SystemLanguageModel, leaving existing SwiftUI views, @Generable output structures, and tool definitions untouched [3]. Apple's own developer materials confirm the framework supports "any provider that conforms to the Language Model protocol," naming Apple Foundation Models, Claude, and Gemini [6].

Gemini is the launch partner, delivered through the Firebase Apple SDK with no separate backend server required and protected by Firebase App Check. It shipped as a preview during WWDC week, with individual developers getting a self-serve key from Google AI Studio and enterprises using the Gemini Enterprise Agent Platform for dedicated quota [2][3].

Apple also added agentic primitives. New Dynamic Profiles let an app swap models, tools, and instructions within a single continuous session, while managing context trimming, summarization, and KV-caching [12][13]. An open-source Foundation Models framework utilities Swift package will ship experimental agentic patterns between OS releases — Apple's framing is that "this field is changing week-to-week" [10]. A new Evaluations framework is positioned as the way to test agentic flows, "going beyond what unit tests alone can catch" [13]. Other additions include multimodal image input, on-device Vision tools (OCR, barcode) callable by the model, semantic search via Core Spotlight, and an fm CLI plus Python SDK [7].

Two more announcements stand out. First, developers in the App Store Small Business Program with fewer than 2 million total first-time downloads get access to next-generation Apple models on Private Cloud Compute at no token cost — MacRumors called this the session's "main announcement" [1][12]. Second, Apple confirmed the framework itself will go open source later in summer 2026 [12].

Separately, Xcode 27 gains agentic coding from Anthropic, Google, and OpenAI, with interactive planning and a canvas for code changes and previews. Agents can validate their own work — writing and running tests, isolating experiments in Playgrounds, and driving the simulator via a new Device Hub — to run autonomously for longer [1].

The models underneath

Apple introduced its third-generation Foundation Models, AFM 3, including on-device AFM 3 Core and a more capable AFM 3 Core Advanced, plus server-side AFM 3 Cloud and AFM 3 Cloud Pro on Private Cloud Compute [11]. In side-by-side human evaluations for general text, AFM 3 Cloud was preferred on 64.7% of prompts versus 8.7% for the 2025 server model, roughly a 36% relative gain in response satisfaction [11]. Notably, AFM 3 Cloud Pro was optimized for NVIDIA GPUs rather than Apple silicon [11].

A distinct framework, Core AI, lets developers run their own weights on Apple silicon's Neural Engine with "zero server dependencies and zero token costs," using a new .aimodel format [10]. Choosing between Core AI, Core ML, and MLX is now a genuine architectural decision.

The nuance worth flagging

Whether the top cloud tier is "Apple's" or "Gemini" is genuinely contested. CNBC reported that AFM 3 Cloud Pro is "similar in quality to Gemini Frontier models" and runs on NVIDIA GPUs in Google's cloud, the product of a collaboration in which Apple reportedly pays Google about $1B per year for access to a Gemini-derived model [16][18]. Apple's own ML Research post avoids the word "Gemini." The honest summary: the branded AFM 3 models are Apple's, but the highest cloud tier is deeply entangled with Google's infrastructure, and the exact boundary remains unconfirmed [11][16].

Why it matters for practitioners

The biggest shift is one Swift API, many backends: write to the framework once and route between an on-device model, Private Cloud Compute, and third-party clouds by swapping a model instance — making portability and fallback routing a first-class, OS-level concern [3][13]. Free Private Cloud Compute inference removes the token bill for small teams shipping server-class features [12]. The Evaluations framework nudges developers to treat prompt and agent regressions as CI-grade concerns [13]. The trade-off is strategic dependency: the most capable cloud path leans on Google, NVIDIA, and a ~$1B annual deal, and the "open" protocol is also a competitive surface where providers will compete for Apple's billions of devices [16][18]. Features were in developer beta at announcement, with consumer rollout slated for fall.