Apple's Core AI, the on-device successor to Core ML

Via Michael Tsai’s roundup, Apple’s WWDC26 Core AI is the framework successor to Core ML: a modern, memory-safe Swift API for running models entirely on device, a set of Python libraries for converting, authoring, and optimizing models, and a coreai-models repository of ready-to-run examples. Models are specialized to the hardware automatically, with ahead-of-time compilation for instant load, and the framework is meant to span everything from compact vision models up to large reasoning models across iPhone, iPad, Mac, and Vision Pro.

The pitch that matters is zero server dependency and zero token cost, with fine-grained control over inference memory and zero-copy data paths. That’s a real platform shift for anyone building on Apple silicon — the kind of foundational plumbing that’s worth a close read now rather than after everyone has built on it. Tsai’s comment thread, as usual, is where the sharp edges and caveats will surface.

What I keep coming back to is the privacy dimension. This is the beginning of genuinely capable local AI — models running entirely on the machine, where the data never has to leave the device to be useful. That changes the default posture from “send it to a server and trust the policy” to “it stays here,” which is a much better place to start from. I’m excited to explore this more as it evolves, and right now it’s evolving at a lightspeed pace.