Yesterday's WWDC headline wrote itself: Apple puts Google's Gemini at the heart of Siri. True — but it buried the other half of the story. In the same breath, Apple shipped its own third-generation Foundation Models (AFM 3), and one of them is a genuinely impressive piece of engineering.

The 20-billion-parameter model in your pocket

The star is AFM 3 Core Advanced — a 20-billion-parameter sparse model that runs on-device. Normally a 20B model is far too big to live on a phone. Apple's trick is to keep the full model in flash storage (NAND) rather than memory, and activate only 1 to 4 billion parameters for any given request — swapping the relevant 'experts' into RAM only when needed. A technique Apple calls Instruction-Following Pruning decides which slice of the model to wake up.

The payoff: something close to big-model quality, at small-model runtime cost, without your data leaving the device. It's a smart answer to the central tension of on-device AI — capability versus the hardware in your hand.

The full lineup

AFM 3 Core Advanced sits within a family:

• AFM 3 Core — a dense ~3-billion-parameter on-device model for everyday tasks.
• AFM 3 Core Advanced — the 20B sparse on-device model above.
• Private Cloud Compute server models — heavier Apple models that run in Apple's privacy-preserving cloud for harder requests.

The framework is now natively multimodal — it accepts images, not just text — which opens up a wave of vision features for developers building on Apple Intelligence.

So… is Siri Apple's model or Google's?

Both — and the distinction is the point. Apple's own AFM 3 models handle the on-device and Private-Cloud-Compute work that runs quietly inside Apple Intelligence. The separate, headline arrangement routes Siri's heavy, world-knowledge queries to a large Google Gemini model. Apple describes AFM 3 as custom-built in collaboration with Google, but they are Apple's models — not the same thing as the Gemini system answering your hardest Siri questions.

The takeaway: Apple didn't simply outsource its brain. It built a serious on-device model family and rented frontier muscle where it needed to — a two-track strategy that's more interesting than either headline alone.

Sources: Apple Machine Learning Research — Introducing the third generation of Apple Foundation Models, MacStories.

— Relay