{"id":69140,"date":"2026-06-12T14:27:27","date_gmt":"2026-06-12T12:27:27","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=69140"},"modified":"2026-06-12T14:28:53","modified_gmt":"2026-06-12T12:28:53","slug":"apple-foundation-models-third-generation-explained","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/allgemein\/apple-foundation-models-dritte-generation-erklaert","title":{"rendered":"Apple's third-generation Foundation Models: How the new AI architecture works"},"content":{"rendered":"

Five models, distributed between iPhones, Apple servers, and third-party cloud hardware: At WWDC 2026, Apple unveiled the third generation of its Foundation Models. For the first time, one of the models does not run on Apple Silicon, but on Nvidia chips in Google's data centers \u2013 a break with a previously ironclad principle.<\/strong><\/p>\n\n\n\n

With the third generation of Apple Foundation Models, internally abbreviated as AFM, Apple is reorganizing the foundation of its entire AI strategy. Instead of a single model, the system now consists of five specialized models that, depending on the task, run on the device, on Apple's own servers, or on third-party infrastructure. This step is the provisional culmination of the realignment of Apple's AI future around Siri and Google's Gemini technology<\/a>, which Apple has been preparing over the past few months. Only now is it becoming concretely clear how this architecture is structured in detail \u2013 and at what point Apple is abandoning its strict \"everything on its own hardware\" principle.<\/p>\n\n\n\n

From the 2024 launch to the Google partnership<\/h2>\n\n\n\n

When Apple first introduced its Foundation Models 2024, the lineup consisted of two components: a language model with around three billion parameters that ran directly on the device, and a larger, server-based language model. The latter was tied to private cloud computing and ran on servers with Apple Silicon chips.<\/p>\n\n\n\n

Private Cloud Compute, or PCC for short, was an ambitious project from the outset. It was designed to provide cloud-based AI capabilities while maintaining the same data privacy guarantees that users are accustomed to from processing directly on their devices. To achieve this, control over all the hardware was essential: PCC ran in Apple's own data centers on Apple Silicon servers, and the data privacy guarantees could be reviewed by independent security researchers.<\/p>\n\n\n\n

However, Apple wasn't progressing quickly enough with its own AI ambitions. As a result, the company partnered with Google and is using its Gemini technology as the backbone of its new AI efforts. Apple emphasizes that this isn't simply Gemini on the iPhone, but rather models specifically modified for Apple based on the Gemini platform<\/a>. Apple presented the results of this collaboration at WWDC 2026.<\/p>\n\n\n\n

Five models, two of them on the device<\/h2>\n\n\n\n

The third generation of AFM comprises five models, distributed across three levels. Two run directly on the device, two on Apple's servers, and one on third-party hardware:<\/p>\n\n\n\n

Model<\/th>Where it's running<\/th>Task<\/th><\/tr><\/thead>
AFM 3 Core<\/td>On the device<\/td>Basic model with three billion parameters, significantly improved quality<\/td><\/tr>
AFM 3 Core Advanced<\/td>On the device<\/td>Most powerful on-device model, natively multimodal, 20 billion parameters<\/td><\/tr>
AFM 3 Cloud<\/td>Apple Silicon Server<\/td>Server workhorse, optimized for speed, efficiency and performance<\/td><\/tr>
ADM 3 Cloud (Image)<\/td>Apple Silicon Server<\/td>Image generation and editing, among other things, powers Image Playground.<\/td><\/tr>
AFM 3 Cloud Pro<\/td>Nvidia GPUs on Google Cloud<\/td>The most demanding tasks, such as agent-based tool use and complex reasoning<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

The "D" in the name AFM 3 Cloud (Image) stands for diffusion, the technology behind image generation. With the exception of AFM 3 Cloud Pro, all models were designed to run on Apple Silicon\u2014either on the device itself or on Apple's servers. AFM 3 Cloud Pro is the exception, running on Nvidia GPUs hosted in Google Cloud. This was made possible because Apple extended its private cloud compute architecture to third-party infrastructure for the first time, reportedly without compromising security and privacy protections. The two most interesting models are AFM 3 Core Advanced and AFM 3 Cloud Pro.<\/p>\n\n\n\n

AFM 3 Core Advanced: 20 billion parameters on the device<\/h2>\n\n\n\n

AFM 3 Core Advanced packs 20 billion parameters into a model that runs directly on the device\u2014a remarkable feat, as most on-device models intended for the general public remain in the low single-digit billions. The model is also natively multimodal, enabling features like more expressive voices and more precise dictation that you'll notice immediately in everyday use. It's unlocked and optimized for Apple's most powerful Apple Silicon systems.<\/p>\n\n\n\n

To make a 20-billion-parameter model function effectively on a device, Apple employs a so-called sparse architecture. Instead of keeping all 20 billion parameters active for every query, as in a dense architecture, the model activates only up to four billion parameters simultaneously, depending on the input. Conceptually, this is similar to the mixture-of-experts approach, but it is based on a technique developed by Apple itself, which the company described a year ago in the study \"Instruction-Following Pruning for Large Language Models.\"<\/p>\n\n\n\n

AFM 3 Cloud Pro and the opening of Private Cloud Compute<\/h2>\n\n\n\n

AFM 3 Cloud Pro is the model that runs on external infrastructure, representing a true break with Apple's previous approach. To deliver Gemini-based peak performance without compromising its data privacy commitment, Apple has, for the first time, extended its PCC architecture to hardware outside its own data centers. The extent of this move is evident in the security measures Apple has implemented in collaboration with Google, which form the same basis as the expansion of Private Cloud Compute to Google Cloud<\/a>.<\/p>\n\n\n\n

Apple explicitly does not rely solely on confidential computing techniques to defend against attacks via privileged access or side channels. Instead, the company includes every component\u2014from firmware and host and guest operating systems to application code\u2014in its trusted base, which is subject to guarantees of verifiable transparency and the absence of privileged access. To combat supply chain attacks, Apple maintains a cryptographically verifiable, extensible directory of all Google Cloud hardware within the PCC network. For particularly sensitive components, software attestation relies on at least two separate trust anchors from independent providers. The processing stack itself follows the same patterns as on Apple Silicon: Incoming data is initially processed in its own isolated process, shared software is recycled after a short time, and cryptographic keys reside in a separate, isolated environment.<\/p>\n\n\n\n

How Apple trained the models<\/h2>\n\n\n\n

According to Apple's research blog, all five models initially shared a common foundation before being specialized for their respective architecture and use cases. This process added multimodal capabilities \u2013 such as understanding audio and images, processing long contexts, and generating high-quality visuals.<\/p>\n\n\n\n

For training, Apple used a mix of publicly available information, third-party licensed or purchased data, open-source data, data collected specifically for studies, and synthetic data. The company emphasizes that neither user data nor interactions were included in the training. Furthermore, web publishers can opt out of the Foundation Models training.<\/p>\n\n\n\n

What the tests show<\/h2>\n\n\n\n
\"Apple
Image: Apple<\/figcaption><\/figure>\n\n\n\n

To evaluate the third generation, Apple relied on extensive human assessments. Internal testers evaluated the models' responses in categories such as following instructions, accuracy, presentation, and image understanding. Where possible, the new models were pitted against their predecessors.<\/p>\n\n\n\n

The comparisons included global English as well as other language groups to demonstrate consistent performance across international variants. For the dictation function, Apple directly compared AFM 3 Core Advanced to the existing dictation system and observed a consistent improvement across seven quality dimensions. For a more in-depth analysis, you can find the complete comparison data on Apple's Machine Learning Research Blog<\/a>.<\/p>\n\n\n\n

Apple's AI architecture between device and third-party cloud<\/h2>\n\n\n\n

The third generation of AFM makes two things clear: Apple is surprisingly venturing to integrate a great deal of processing directly into the device with AFM 3 Core Advanced, while pragmatically relying on Google's Gemini technology and third-party hardware for the most demanding tasks. The real challenge lies less in the models themselves than in the attempt to uphold the privacy promise of Private Cloud Compute even when the computation takes place in a third-party data center. Whether this balancing act will live up to Apple's promises in practice will become clear once the new Siri and the other features are widely rolled out.<\/p>\n\n\n\n

The best products for you: Our Amazon Storefront<\/a> offers a wide selection of accessories, including for HomeKit. (Image: Apple)<\/p>\n\n\n\n