Apple's intelligence is a set of composable frameworks, not a single feature. This page maps the components — what each one is, and where it meets the work of building and governing AI on the platform. Apple's capabilities are stated plainly. Where I add a practitioner's read, it's marked as mine.
The native Swift API for Apple's models. It takes multimodal prompts — text and images — and integrates the Vision framework, so the model can call purpose-built tools like on-device OCR and barcode reading. It can also call server models, including Claude and Gemini, for tool calling and guided generation; any provider can ship a Swift package conforming to the language-model protocol, which makes the underlying model swappable behind one API. Developers under a first-time-download threshold can run Apple's models on Private Cloud Compute at no cloud-API cost. The framework is open source, and the same Swift APIs run server-side. Around it sit an Evaluations framework, a Foundation Models instrument, an FM command-line tool, a Python SDK, and a RAG tool backed by Core Spotlight.
Apple's Foundation Models are built on the technology behind Google's Gemini family, then adapted to run on device and on Private Cloud Compute.
A declarative API for configuring a model session. In a result-builder syntax, the active profile recomputes every model turn and always resolves to one profile driving the session — sharing a single continuous transcript while the app swaps model, instructions, tools, and sub-agents based on state. Apple positions it as a building block for agents and skills.
A framework for bringing and running your own models on device: a memory-safe Swift API, custom GPU kernels, model specialization, PyTorch-to-Core-AI conversion, ahead-of-time compilation, and a tensor-level visual debugger that traces back to the original Python source. No server dependency; it scales from compact vision models to multi-billion-parameter LLMs.
The Mac's transition to Apple silicon is complete, so the whole ecosystem is one architecture. Apps can ship as Apple-silicon-only binaries on the Mac App Store.
In the language: anyAppleOS availability, scoped warning suppression, async in defer, and real relief for the “unable to type-check in reasonable time” error. In the system: parts of the OS kernel are written in Swift, alongside WebKit, the QUIC transport layer — open-sourced through SwiftNIO — and the font engine.
Reorder and swipe actions in any container, fuller text selection, faster layout on nested stacks, @State as a macro with lazy initialization, HTTP-cached AsyncImage, finer toolbar control, document infrastructure with first-class URL access, and a Spatial Preview framework that streams 3D to Apple Vision Pro.
How an app becomes a participant in the system rather than an island. App Intents expose an app's content and actions to Siri, Spotlight, and Shortcuts; entity schemas contribute that content to the Spotlight semantic index; and View Annotations let a person refer to what is on screen in conversation. The app's own capabilities become things the system intelligence can find and call.
Refinements to content diffusion and edge contrast, plus a user tint slider from clear to fully tinted. The opt-out for the old design is going away — recompiling adopts Liquid Glass automatically.
iOS apps resize in iPhone Mirroring and on iPad, opted in on an SDK rebuild. The target shifts from fixed devices and orientations to a dynamic range of sizes.
Agents run through the toolchain: a plan step with rendered diagrams, then agents that build, run tests, drive the simulator, check Previews across variants, localize in context, and triage and fix top crashes from the Organizer. Built-in “specialists” — SwiftUI, accessibility, sizing, testing, performance — ship as a corpus of skills, docs, and MCP tools, and you can add your own.
Plugins are the extensibility format: skills as markdown, MCP tools, and the Agent Client Protocol to bring any agent. Built-in integrations cover Anthropic, OpenAI, and Google.
Apple's array framework for training, fine-tuning, and local inference on Apple silicon — with Metal 4, the GPU Neural Accelerators, and multi-Mac training over RDMA and Thunderbolt.
A framework for testing and validating prompts and model behavior, paired with the Foundation Models instrument for visualizing and debugging what the model does in your app.
Reality Composer Pro 3, the Game Porting Toolkit, and Metal command-line tools extend the same agent-driven workflow to 3D and games.
Each becomes its own track on this site, where the practitioner's read above is carried all the way through. Some are live; others are forming.
What is happening, and what deserves attention?
Turning raw signal into a current, attributed operating picture. The Dispatch.
Enter →What may the model know, and how is it grounded?
Grounding answers in retrieved, attributed data — the private semantic index, personal context — not the model's guesses.
Forming →Where should intelligence run?
Governed intelligence on the device, off the network.
Enter →When is reaching past the device justified?
Escalation to Private Cloud Compute and server models — deliberately, not by default.
Forming →How does an app take part in the system?
Exposing an app's content and actions to Siri, Spotlight, and Shortcuts — App Intents, under control.
Forming → The keystoneWho has authority when it acts?
Deterministic authority over what an agent is allowed to do. SwiftVector and Vigil.
Enter →How does building change when agents take part?
Building with coding agents without surrendering authorship.
Forming →How does a person stay formed enough to command it?
Forming the judgment to stay in command when the machine is helpful, wrong, or silent.
Enter →That's the stack as it stands: three layers, a developer surface that now reaches the server, a consumer product that stayed on the edge. The next question — who has authority when it acts — is where the disciplines begin.