These Apple third generation models include five distinct models designed for both on-device and server-based AI tasks. Two of these models run directly on your device, while three operate on servers through Private Cloud Compute. This split allows for efficient machine learning without compromising privacy or performance.
Known collectively as Apple Foundation Models, this third-generation AI lineup is custom-built for on-device machine learning and server-based processing. The on-device models handle lightweight, real-time tasks, while the server models tackle more complex requests. This practical approach ensures you get reliable AI capabilities across your Apple devices.
The Sparse Architecture of AFM 3 Core Advanced
While the previous section explained how Apple splits workloads between on-device and server models, the real magic happens in the flagship on-device model: AFM 3 Core Advanced. This 20-billion-parameter model uses a sparse architecture — specifically a sparse mixture of experts — so that it only activates 1 to 4 billion parameters per request. That dynamic parameter activation is what keeps the model both powerful and efficient enough to run directly on your iPhone or iPad. Rather than firing up all 20 billion parameters for every task, the model selects a predetermined number of active experts tailored to exactly what you’re asking. This is a defining characteristic of Apple’s Apple third generation models: they prioritize practical, on-device performance without compromising capability.
How does Apple achieve this? The secret lies in Instruction-Following Pruning (IFP), a technique developed by Apple researchers. IFP trims away unnecessary pathways during inference, ensuring only the most relevant parameters are used for a given request. The full 20-billion-parameter model stays stored in flash memory, and the system selectively loads the required sparse experts into DRAM when needed — a process known as flash memory inference. This combination means you get fast, responsive AI features (like smarter Siri or real-time photo analysis) without draining your battery or bogging down the processor. For everyday use, that sparse architecture is what makes the on-device experience feel both effortless and intelligent.
On-Device AI: AFM 3 Core and Core Advanced
That blend of speed and battery smarts comes from a thoughtful two-model approach in the Apple third generation models. Instead of one-size-fits-all, Apple gives you two on-device options: a lightweight 3-billion-parameter dense model called AFM 3 Core, and a more powerful 20-billion-parameter sparse model dubbed AFM 3 Core Advanced. The dense model is built for efficiency — it runs common tasks like typing suggestions, smart photo tagging, and quick voice commands with minimal power draw. That means your iPhone or iPad handles these jobs locally, without reaching for the cloud. The sparse model, by contrast, activates only a fraction of its parameters at any given moment, so it can handle heavier lifting — say, composing a longer email draft or summarizing a document — without overwhelming your device’s resources.
AFM 3 Core: Efficient Dense Model
The AFM 3 Core is exactly what you want for everyday edge AI. At 3 billion parameters, it’s compact enough to fit comfortably on a phone but still capable of understanding natural language and generating responses. Dense models use all their parameters for every task, which here means reliable, consistent performance for straightforward on-device inference. So when you ask Siri to set a timer or open a specific playlist, the dense model handles it instantly. It’s also the workhorse behind real-time features like live text translation and camera scene recognition — actions you expect to happen in a split second. Because it runs entirely on-device, your privacy stays protected: nothing leaves your phone.
AFM 3 Core Advanced: Sparse Powerhouse
For more demanding jobs, the 20-billion-parameter sparse model steps in. Sparse activation means only the relevant parts of the network wake up for each request. That makes it possible to run a much larger model on mobile hardware without the usual drain on battery or memory. AFM 3 Core Advanced can handle complex reasoning tasks, like helping you rewrite a paragraph in a different tone or generating a short poem from a few keywords. You’ll notice it in tools like the Smart Compose feature in Messages or the Privacy Review summarizer in Settings. The model is always there when you need it, but it sips power rather than gulps it. Together, these two models give you the best of both worlds in mobile machine learning: everyday speed from the dense engine, and deep intelligence from the sparse powerhouse.
Server-Based Models and Private Cloud Compute
When a task is too demanding for your device, Apple third generation models shift the workload to the cloud — but in a way that respects your privacy. The company’s three server-based models run on Private Cloud Compute, a system designed to ensure your data is never stored or shared. This approach to confidential computing means your requests are processed and then immediately discarded. You get the power of server-side AI without the usual privacy trade-offs.
These models handle complex tasks that go beyond on-device capabilities. The lineup includes AFM 3 Cloud, ADM 3 Cloud (for image processing), and AFM 3 Cloud Pro. Each one is built for a specific purpose, whether it’s language understanding, image generation, or heavy-duty computation. The key difference from typical cloud AI is that Private Cloud Compute acts as a secure bridge — your data never leaves the encrypted pipeline. So when you need a quick answer or a creative output that your phone can’t handle alone, the server models step in without compromising data privacy.
AFM 3 Cloud Pro: Agentic Tool Use and Complex Reasoning
That same privacy-first approach extends further with the most advanced server model yet. While the standard cloud models handle creative and analytical tasks, the Apple third generation models push into a new territory with AFM 3 Cloud Pro. This model is built for agentic AI — meaning it can act autonomously, use tools, and perform multi-step reasoning on your behalf. Instead of just answering a single question, it can plan a sequence of actions, call external APIs, and synthesize results. Think of it as a capable assistant that can book appointments, cross-check data from multiple sources, or run logical chains that involve several decisions.
Agentic Capabilities
What sets this model apart is its ability to handle complex reasoning tasks that require planning and tool use. For example, if you ask it to compare flight options, check your calendar, and suggest a meeting time, it can break that down into steps, use the right tools, and return a coherent answer. This moves beyond simple Q&A into genuine task completion.
Privacy-Preserving GPU Acceleration
A notable technical step is how AFM 3 Cloud Pro extends Private Cloud Compute to run on NVIDIA GPUs in Google Cloud. This means the heavy lifting for agentic tasks happens on powerful hardware, but your data remains encrypted and inaccessible to the cloud provider. The model itself runs in a secure enclave, so you get the speed of GPU acceleration without sacrificing privacy. For anyone who needs reliable, autonomous AI assistance without handing over their personal information, this is a practical leap forward in what cloud AI can do responsibly.
Collaboration with Google: Custom-Built Foundation Models
Apple’s third-generation models aren’t just designed in Cupertino — they’re built hand-in-hand with Google. This Apple Google partnership means the entire AFM family of five foundation models is custom-built, not off-the-shelf. By combining Apple’s AI research expertise with Google Cloud infrastructure and NVIDIA GPUs, the collaboration delivers custom AI models that are both powerful and privacy-respecting. That’s a rare combination in the cloud AI space.
So how does it work? Apple’s models run on Google Cloud, but with a twist: the infrastructure is optimized to process requests without storing or exposing user data. NVIDIA collaboration brings hardware acceleration that keeps inference fast, while Apple’s proprietary privacy layers ensure that even when your data leaves your device, it stays encrypted and anonymous. For you, that means server-side AI tasks — like complex image recognition or language understanding — happen quickly and securely. It’s a practical example of how custom-built foundation models can scale efficiently without compromising the trust you expect from Apple.
Frequently Asked Questions
How does AFM 3 Core Advanced’s sparse architecture achieve selective activation of only a fraction of its total parameters?
The model uses a learned routing mechanism that identifies the subset of parameters most relevant to your current task. Only that small group is activated, dramatically reducing power and memory usage. This design keeps on‑device processing fast and efficient while still delivering high‑quality results.
What specific on‑device features does AFM 3 Core Advanced enable that the previous generation could not?
With Apple third generation models, you get real‑time language understanding for tasks like advanced summarization and context‑aware smart replies that run entirely on your device. These models also handle more complex natural language interactions directly in apps, without ever needing a cloud connection—something the previous generation couldn’t do locally.
What guarantees does Private Cloud Compute provide to ensure user data is never stored or shared?
Private Cloud Compute uses cryptographic attestation to verify that only Apple‑approved software runs on its servers. Your data is processed solely for your immediate request and is never logged, retained, or accessible by Apple. The infrastructure is designed to prevent any permanent storage, so your privacy is protected at every step.
