Apple Foundation Models are the model family behind Apple Intelligence. On June 8, 2026, Apple introduced AFM 3, its third generation of foundation models, and the important shift is bigger than a Siri update: Apple is turning foundation models into an operating-system and developer-platform layer.
The new AFM 3 family spans on-device models and server-based models running on Private Cloud Compute. Apple also expanded the Foundation Models framework so developers can build multimodal, agentic, and model-flexible AI experiences inside apps.
That makes this one of the most important AI platform announcements of the past 24 hours. Apple is not trying to win only by shipping a standalone chatbot. It is trying to make AI available where users already work: inside iOS, macOS, Siri, Photos, Xcode, and third-party apps.
What Apple Announced
Apple published a technical overview of the third generation of Apple Foundation Models on June 8, 2026. The new family includes five models:
- AFM 3 Core: a 3-billion-parameter dense on-device model.
- AFM 3 Core Advanced: Apple's more powerful on-device model, described as natively multimodal and optimized for capable Apple silicon systems.
- AFM 3 Cloud: a server-side model for Private Cloud Compute.
- ADM 3 Cloud: an image generation and editing model for features such as Image Playground and photo editing.
- AFM 3 Cloud Pro: Apple's most capable server-based model for heavier work such as complex reasoning and agentic tool use.
Apple says the AFM 3 family powers a new version of Siri, advanced photo-editing tools, Image Playground updates, expressive voices, and other Apple Intelligence experiences. The company also says these models run either on-device or on Private Cloud Compute, not on a generic third-party API path.
The developer story matters just as much. In its WWDC26 Apple Intelligence guide, Apple says the Foundation Models framework now gives developers access to the same on-device model that powers Apple Intelligence, and it can work with Apple Foundation Models, cloud models such as Claude and Gemini, or other providers that conform to Apple's Language Model protocol.
What Is AFM 3?
AFM 3 is Apple's third generation of Apple Foundation Models. It is not one model. It is a model family designed for different execution environments and tasks.
The most interesting technical detail is AFM 3 Core Advanced. Apple describes it as a 20-billion-parameter sparse model that activates only 1 to 4 billion parameters depending on the request. The practical goal is clear: Apple wants more model capability on consumer hardware without loading the whole model into active memory for every task.
Apple says AFM 3 Core Advanced stores the full model in flash memory and loads selected expert weights into DRAM based on the prompt. That design lets the model scale beyond the normal memory limits of a phone or laptop while keeping the active computation smaller.
This is the kind of architecture that fits Apple's priorities:
- Run useful AI locally when possible.
- Keep personal context close to the device.
- Use private cloud models only when the task needs more capacity.
- Expose the model through platform APIs instead of making every app integrate a separate AI backend.
The result is a different model strategy from the usual cloud API race. Apple is optimizing for device integration, privacy claims, power efficiency, and app distribution, not just public benchmark leadership.
How the Foundation Models Framework Changes App Development
The Foundation Models framework is the developer access point. It lets app developers build features around Apple Intelligence models through native Swift APIs.
The WWDC26 update expands that surface in several important ways:
- Multimodal prompts: apps can pass images alongside text.
- On-device tool use: Vision framework tools such as OCR and barcode readers can be called locally.
- Dynamic Profiles: apps can change models, tools, and instructions during a session.
- Provider flexibility: developers can work with Apple Foundation Models, cloud models like Claude and Gemini, or other providers that conform to Apple's protocol.
- Evaluations: Apple is pointing developers to tools for testing AI behavior beyond traditional unit tests.
That last point matters. AI features fail differently from normal software. They can be correct in one context and unreliable in another. A platform-level evaluation workflow is a sign that Apple wants developers to treat AI features as testable product behavior, not just prompt experiments.
Why On-Device Plus Private Cloud Matters
Most AI apps start with a simple architecture: send the user request to a cloud model, receive the answer, display it in the app.
Apple is pushing a different pattern:
- Use on-device models for private, fast, everyday intelligence.
- Use Private Cloud Compute for heavier reasoning, generation, and agentic work.
- Let apps connect through native frameworks and system entities.
- Use App Intents and system context so Siri and Apple Intelligence can act inside apps.
For users, this can mean less copying information into chat windows. For developers, it means AI features can sit closer to the app's own data, UI, and workflow. For enterprises, it gives Apple a clearer privacy and governance argument than a generic cloud LLM integration.
The tradeoff is platform dependency. If your AI roadmap depends heavily on Apple-only APIs, you may get a smoother experience on iOS and macOS, but you also need a separate strategy for Android, Windows, and the web.
Why This Is a Platform Strategy, Not Just a Model Launch
The most important phrase in this announcement is not "20 billion parameters." It is "framework."
Models create capability. Frameworks create distribution.
By putting Foundation Models behind native APIs, Apple can make AI part of the default app-building environment. Developers do not need to start with billing for a separate model provider, a custom inference service, a vector database, and a pile of privacy reviews just to test whether an AI feature belongs in an app.
Apple is also using economics as part of the strategy. Its WWDC26 guide says eligible App Store Small Business Program developers with fewer than 2 million total first-time App Store downloads can access next-generation Apple Foundation Models on Private Cloud Compute at no cloud API cost.
That does not make Apple AI "free" in every context. It is a specific program with eligibility constraints. But it does lower the early experimentation barrier for the exact group Apple wants to keep building native apps: small teams and indie developers.
Practical Implications for Developers
If you build on Apple platforms, AFM 3 changes the AI feature checklist.
First, decide what should run locally. Features involving personal context, quick text transformations, image understanding, local search, or lightweight classification may be better candidates for on-device models than a cloud call.
Second, design fallback paths. A feature may need to switch between on-device Apple models, Private Cloud Compute, and a third-party provider depending on task complexity, user settings, device class, region, or cost.
Third, add evaluations early. Prompt behavior can drift when instructions, tools, and model profiles change. Apple's Evaluations framework is a signal that developers should test agentic flows like product behavior, not one-off demos.
Fourth, watch App Intents. If Siri AI can understand app entities and trigger app capabilities through natural language, the discoverability surface for apps changes. Users may not always open your app first. They may ask the system to act through it.
Limitations and Open Questions
There are several reasons to stay careful.
Apple's model quality claims are first-party. They are useful as product context, but they are not a substitute for independent benchmarks or real app testing.
Availability details also matter. Device support, region restrictions, language coverage, beta behavior, and production limits can all affect whether a feature is viable for your user base.
The framework abstraction is promising, but abstraction does not remove operational work. Developers still need to handle latency, model fallback, user consent, privacy expectations, evaluation, logging policy, and error states.
There is also a portability risk. Apple-native AI features may be the fastest path to a polished iPhone or Mac experience, but cross-platform products should avoid building their entire intelligence layer around one operating system's model stack.
What Builders Should Watch Next
The next signals are practical, not theatrical:
- Which devices can run AFM 3 Core Advanced well?
- How reliable are multimodal prompts in real apps?
- How easy is it to swap between Apple models and third-party providers?
- What are the production limits for Private Cloud Compute access?
- How well do App Intents and Siri AI drive users back into third-party apps?
- Do developers publish real benchmarks for latency, cost, and quality?
If those answers are strong, AFM 3 could make Apple devices a serious default environment for private, embedded AI experiences.
If those answers are weak, the announcement will still matter, but more as a platform signal than as an immediate developer unlock.
Conclusion
Apple AFM 3 is not just another model family. It is Apple's attempt to make foundation models part of the operating system and the app-development stack.
The big idea is simple: AI should not always feel like leaving your workflow to ask a chatbot. On Apple platforms, the model may increasingly sit inside the app, inside the system, and inside the user's normal device context.
That is why this announcement matters. The next phase of foundation models is not only about larger models in the cloud. It is about where models live, how developers reach them, what privacy guarantees surround them, and how naturally they fit into daily software.
FAQ
What are Apple Foundation Models?
Apple Foundation Models are the generative models that power Apple Intelligence experiences. The third generation, AFM 3, includes on-device models and Private Cloud Compute models for tasks such as Siri, image generation, editing, expressive voice, and agentic tool use.
What is AFM 3 Core Advanced?
AFM 3 Core Advanced is Apple's more powerful on-device model in the AFM 3 family. Apple describes it as a 20-billion-parameter sparse model that activates 1 to 4 billion parameters depending on the request, allowing more capability on Apple silicon devices.
Do Apple Foundation Models run on-device?
Some do. AFM 3 Core and AFM 3 Core Advanced are on-device models. Apple also uses server-based AFM 3 models through Private Cloud Compute for heavier tasks.
Can developers use Apple Foundation Models in their apps?
Yes, Apple says the Foundation Models framework gives developers native Swift API access to Apple Intelligence's on-device model and broader model options. Developers should confirm exact availability and beta behavior in current Apple Developer documentation.
How is Apple Foundation Models different from using a cloud LLM API?
Apple's approach is more platform-native. It combines on-device models, Private Cloud Compute, App Intents, Siri AI, and native developer frameworks. A cloud LLM API can be more cross-platform and model-flexible, but it usually requires a separate backend, billing, and privacy architecture.