Foundation Models is Awesome

This year I was lucky enough to attend WWDC25 in person, and what an event it was! There were plenty of big announcements-new designs, the debut of Apple Intelligence, and, my personal favourite, Foundation Models. If you haven’t heard, Foundation Models let you tap into Apple’s large language models (LLMs) right on any device with Apple Intelligence enabled.

There’s been a lot of chatter online suggesting these models aren’t anything special, just because they run on-device and power Apple Intelligence the same feature that got some heat for not living up to early hype. But after digging into the technical paper and playing with the SDK, I’ve realised the 3-billion-parameter on-device model is actually something unique.

First off, Apple didn’t set out to build a “one-size-fits-all” chatbot like GPT-4o or Gemini Flash. Instead, the 3B model is intentionally designed for tasks that truly matter on-device: things like text extraction, summarisation, and generating structured outputs. Apple’s paper makes it clear they’re not chasing benchmarks or flashy general-purpose chatbots. They’re building a model that’s genuinely useful for everyday tasks.

What really stands out is how the model is optimised for Apple silicon, making it fast and efficient. The architecture is clever: it splits the model into two blocks, sharing memory between them to reduce resource use and speed up responses. That means you get near-instant answers without draining your battery or sending your data to the cloud.

Another thing I love is the model’s focus on structured output and tool calling. Using the new Foundation Models framework, you can generate entire Swift data structures with just a few lines of code, thanks to the ‎⁠@Generable⁠ macro. The SDK makes it super easy to get reliable, type-safe resultsno more wrestling with messy text outputs. And with tool calling, you can extend the model’s abilities by letting it interact with custom tools or services you build.

Apple’s approach is a breath of fresh air. Most users don’t want to write code or pen the next great novel with their on-device AIthey want help with real, practical tasks. Apple’s not trying to impress with leaderboard scores; they’re aiming to deliver AI that quietly and reliably makes life easier.

As an iOS developer, I’m thrilled that Apple is opening up these models to us. After experimenting with the SDK, I can honestly say it’s my new favourite framework. The “snapshot” approach to output where you get the whole structured result at once, not just a stream of tokens is a game changer. Watching UIs update live as the model generates output is honestly amazing.

Here’s what this means for Tetrix: we’re making a strong shift toward using on-device models wherever possible (keeping in mind that foundation models are only available on iOS 26+ and Apple Intelligence-enabled devices). We’ll still use the cloud for some AI tasks, but now that there’s a simple on-device system, we can plug things like health data into Tetrix Context to keep everything secure. Also, if you turn on Advanced Security in settings, we’ll only use on-device models assuming your device supports Apple Intelligence.

In short, Apple’s Foundation Models feel very “Apple”thoughtful, focused, and designed to empower both users and developers. If you haven’t checked out the introduction video or played with the framework yet, I highly recommend it. This is the kind of AI that just works, right out of the box.

Wish list for the SDK:

• Support for the private cloud compute model within the SDK
• The ability to have foundation models work on all devices on iOS 26. This would mean that rather than opting for the local model the SDK would just default to the cloud model.
• The ability to generate types that conform to identifiable while using snapshot streaming. Without using funny workarounds.

Resources:

• https://developer.apple.com/documentation/foundationmodels
• https://machinelearning.apple.com/research/apple-foundation-models-2025-updates