The headline of the Apple developer conference was a siri reborn. The most interesting story is below: the AI models Apple built to run it, one of which is too big to fit in an iPhone’s memory, but runs on the device anyway.
In a technical post published in conjunction with WWDC, Apple detailed the third generation of its Apple Foundation models, a family of five models that it describes as “tailored in collaboration with Google.”
Two run on the device: AFM 3 Core, a 3 billion-parameter model for everyday tasks, and AFM 3 Core Advanced, its most powerful on-device model. Three more run in the cloud: AFM 3 Cloud, a server workhorse; ADM 3 Cloud, an image model behind Image Playground and Genmoji; and AFM 3 Cloud Pro, the heavyweight built for agent tools and complex reasoning.
Intelligent engineering is in Core Advanced.
The 💜 of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
It’s a native multimodal model with 20 billion parameters, the kind of size you’d typically find in a data center, not a phone. Apple’s trick is to keep the entire model in flash storage instead of the much smaller working memory. Using a technique its researchers call instruction-trace pruning, the model makes routing decisions once per message, loading only a small set of “expert” parameters into memory, between 1 billion and 4 billion at a time, while maintaining an always-on core of shared experts.
That allows Apple to scale the model “well beyond the traditional limits of DRAM,” he says, and powers more expressive voices and clearer dictations in this year’s software.
The cloud models are based on Apple’s Private Cloud Compute, which the company says prevents user data from being stored or shared with anyone, including Apple. For the high-end Cloud Pro model, Apple worked with Google and NVIDIA extend that privacy architecture to Nvidia GPUs in Google Cloud.
That partnership with Google is the detail worth untangling. Coverage of the keynote variously suggested that Apple’s models were “Gemini distillates” or contained no Google technology at all.
The technical post falls in the middle: the AFM family is Apple’s own, “custom-built in collaboration with Google” and trained on Google’s cloud TPUs, while the heavier reasoning behind the new Siri is supposedly based on a large custom model from Google. In short, the models are from Apple, the muscle and much of the infrastructure are from Google.
For developers, the most important change is the framework of the basic models.
Apps can leverage the device’s model directly, and this year Apple added a model abstraction layer that allows developers to swap in third-party models like Anthropic’s Claude or Google’s Gemini without rewriting their code, while iOS 27 will let users set a rival assistant as the default. It’s an unusually open stance on Apple’s part, even if Apple Intelligence itself not yet coming to the EU on the same timeline.
The usual caveat applies to numbers. Apple’s post is littered with flattering comparisons—AFM 3 Cloud was preferred over last year’s model in 64.7 percent of requests, expressive voices scored 4.15 on a 5-point opinion scale versus 3.87 for the previous system—but these are Apple’s human evaluations, not independent benchmarks, and the models are still in beta.
A more complete technical report is promised later this summer.
Still, after two years of teasing over an assistant that didn’t work, this is Apple’s clearest argument that the plumbing is finally real: a small, private model for everyday use, larger models locked inside their own cloud for the hard stuff, and Google’s frontier muscle where Apple still can’t compete alone.
Whether it stays off Apple’s lists is the next test.
