Monday’s Worldwide Developers Conference wasn’t about fireworks. It was about infrastructure.
Apple finally pulled back the curtain on Apple Intelligence, the upgrade everyone has been whispering about for months. The biggest surprise? It isn’t built in a vacuum. Apple teamed up with Google to train five new foundational AI models. They are the unseen engines.
You won’t talk to “AFM 3” directly. You won’t see it as an app icon. But it powers the Siri you will use every day. And yes, there are serious implications for how your data moves through Google Cloud and Nvidia servers.
Here is exactly how these Apple and Google AI models are structured, why they matter, and where the privacy line actually sits.
The Five Apple Foundation Models
Forget catchy names. This is engineering. Apple released a family of models simply labeled by capability and scale.
They serve different roles:
- AFM 3 Core: A 3-billion parameter model. It lives on your device. Think iPhone. Think Mac. It’s fast, local, and small.
- AFM 3 Core Advanced: This is Apple’s heavyweight local player. A 20-billion parameter multimodal model. Apple calls it the “most powerful on-device model” available.
- AFM 3 Cloud: Runs on servers, not silicon you carry. Heavy lifting happens here.
- AFM 3 Cloud (image): Specialized. It drives the new AI photo editing and Image Playground features.
- AFM 3 Cloud Pro: The big gun. Designed for “demanding use cases,” like agentic tool use and deep, complex reasoning.
These models are the road. Your interface is the car. And the car is Siri.
Craig Federighi, Apple’s senior vice president of software engineering, put it bluntly:
“We believe that truly helpful AI must be center around you and your needs.”
Why On-Device AI Is The Main Selling Point
Apple is betting everything on keeping data close.
Most of these features lean hard toward on-device processing. The smaller models—like the AFM 3 Core —run entirely inside your hardware. No internet connection needed. No cloud round-trip.
This is where the Apple and Google AI models diverge from competitors.
Rivals often default to cloud-based processing. Apple wants the computation to happen in the palm of your hand. The benefits? Latency drops to near zero. Privacy increases instantly.
But when the task is too complex? Like writing a long report or generating high-res images from scratch? The system hands it off to the cloud models. Specifically, the AFM 3 Cloud series.
Here is the friction point.
The AFM 3 Cloud Pro runs through Google Cloud infrastructure. It uses Nvidia GPUs. Technically, your data leaves Apple’s controlled environment to hit a server powered by a rival’s cloud network.
Apple insists this is still private. They call it Private Cloud Compute.
- They claim chat logs and data aren’t stored.
- The data is processed in isolation.
- It’s deleted immediately after the response is generated.
The average user might shrug. “I don’t care where it processes as long as it works.”
But for enterprise? For lawyers, doctors, journalists? These details are everything. Does Google have a backdoor? Does Apple? Or is the architecture truly air-gapped during inference? Apple says yes to privacy. Google has signed off. Nvidia provided the iron. Trust the ecosystem or trust nothing?
What’s Actually New in iOS 19 and macOS Tahoe
Wait. iOS 19?
The article mentions macOS Golden Gate (likely referring to the Tahoe or Sequoia branding cycle adjustments in early leaks) and hints at iOS 17/18/19 iterations. The source text mentions “upcoming iOS 27” which appears to be a typographical error for iOS 18 (the current cycle) or a futuristic placeholder. Given the context of WWDC, we are discussing the current iOS 18 and macOS Sequoia/Golden Gate ecosystem. Let’s stick to what we know is real.
Siri is the face of this change.
It’s no longer the clumsy assistant you used to ask “What’s the weather?”
The new Apple Intelligence allows Siri to understand context across your device. It reads your notes. It understands your calendar. It knows that “call my mom” might actually mean “text her I’m running late.”
This requires multimodal understanding.
AFM 3 Core Advanced sees and hears. It connects visual data (a screenshot of a receipt) with linguistic queries (“how much did I spend on lunch?”).
Before? Siri couldn’t do that. Not really. Now it can.
Invisible AI Is the Strategy
Google builds flashy tools. Nano Banana. ChatGPT interfaces. Claude Code. Viral moments.
Apple wants no viral moments.
They want AI to disappear.
Francisco Jeronimo, a VP at analyst firm IDC, put it best in an email to reporters:
“The impact could be significant… If Apple makes AI feel natural, private, and useful… it could redefine what consumers expect.”
Apple doesn’t want you thinking about the AI. They want you using your phone and having it just work.
Is that lazy? Or is it mature?
Compare this to Google. Google is pushing you to learn new tools. Apple is hiding the tools inside the ones you already own.
For many users, that is less intimidating. Less cognitive load. You don’t need to prompt-engineer. You just speak. Or look at a photo.
But the cost is transparency. You rarely see what is happening. When a model switches from on-device to cloud-based, you don’t get a notification. It just feels faster or slower.
The Developer Side: Core ML Meets Core AI
Developers have a new toolkit: Core ML.
No, not just machine learning frameworks. Apple is rolling out specific APIs that allow developers to build agents using these foundation models.
You can build apps that use the AFM 3 models natively.
If you want an app that edits photos without leaving your phone, you tap into AFM 3 Cloud (IMAGE) or the local Core Advanced. If you want heavy data processing, you leverage the cloud Pro models.
It’s a tiered system.
It gives developers the choice:
- Opt for speed and privacy? Use the 3-billion or 20-billion on-device models.
- Opt for raw power and complex logic? Use the cloud Pro.
But you pay a toll. The cloud models cost compute time. And remember—the Apple and Google AI models partnership means some of that compute lives outside Cupertino.
Who Wins?
If Apple’s prediction is correct, this will feel “invisible.”
We might never hear the words “AI model” again. We’ll just say, “Hey Siri, fix this.”
And if the AFM 3 Core does the heavy lifting on-device? Perfect.
If it hands off to AFM 3 Cloud Pro? You’ll never know. Unless you check the data usage stats. Or read the privacy report.
Most people won’t.
So the question isn’t which model is better. The question is: Do you trust Apple to handle the Google hand-off correctly?
We will find out when iOS hits public devices. Until then, we stare at slides. We guess at parameter counts. And we wait.
Why rush? The tech will be here whether you look at it or not.
