← Back to feed

Apple announced five third-generation Foundation Models at WWDC26, replacing the architecture behind Apple Intelligence with a lineup that spans from a 3-billion-parameter on-device model to a server-side tier running on Nvidia hardware inside Google Cloud. The announcement corrects the pre-WWDC reporting, which characterised iOS 27’s Siri as running on Google Gemini directly. The reality is more nuanced and architecturally more interesting.

The Five Models

On-device:

• AFM 3 Core: Next generation of the 3B-parameter dense on-device model. Handles everyday Apple Intelligence tasks locally.
• AFM 3 Core Advanced: Apple’s most capable on-device model. 20 billion total parameters, but runs a sparse architecture that activates only 1 to 4 billion parameters per request. Natively multimodal. Requires Apple’s most powerful silicon.

Private Cloud Compute (Apple Silicon servers):

• AFM 3 Cloud: Server-side workhorse for speed, efficiency, and performance. Runs on Apple Silicon in Apple’s Private Cloud Compute infrastructure.
• ADM 3 Cloud (Image): Diffusion model for image generation and editing. Powers Image Playground and advanced photo-editing tools. Runs on Private Cloud Compute.

Google Cloud (Nvidia GPUs):

• AFM 3 Cloud Pro: Apple’s most capable server model, used for agentic tool use and complex reasoning. Runs on Nvidia GPUs hosted in Google Cloud — the one model in the lineup that is not on Apple-owned infrastructure.

The Gemini Question

Pre-WWDC reporting from Bloomberg said Siri would be rebuilt on a Gemini foundation at $1 billion per year in licensing fees. Apple’s own technical disclosures tell a different story.

Apple says all five AFM-3 models “shared a common initial foundation before specializing.” For the four Apple Silicon models, Apple describes the training process as: “proprietary data with reinforcement learning and refined using outputs from Gemini frontier models.” That language, confirmed by Craig Federighi in session materials, means Apple trained its own weights, then used Gemini outputs as a signal during reinforcement learning — a form of knowledge distillation, not licensing. At inference time, no Gemini API is called. Apple models run Apple weights.

The exception is AFM 3 Cloud Pro, which runs on Nvidia-and-Google infrastructure. Apple’s Private Cloud Compute documentation does not cover this model. The architectural split — four models on Apple Silicon, one on Nvidia/Google — is the clearest signal of where Apple’s infrastructure capability hits its ceiling at the frontier tier.

AFM 3 Core Advanced: The Flash Architecture

The most technically notable model is AFM 3 Core Advanced, the 20B on-device model. The central problem with running a 20B model on a phone or Mac is memory: standard models require all weights in DRAM. Apple’s answer is Instruction-Following Pruning (IFP), developed internally since at least January 2025.

IFP stores the full model in NAND flash memory. On each prompt, a lightweight dense block makes routing decisions upfront — selecting which expert weights will be needed — and loads only those into DRAM before generation begins. A mix of “always-active shared experts” and “routed experts” means DRAM holds a fraction of the total parameter count at any moment. The model never swaps weights token-by-token, which would be too slow given NAND-to-DRAM bandwidth limits.

The result: a 20B total parameter model runs in the DRAM footprint of a 1-4B model, depending on the task. Apple claims a “step-change in on-device quality” relative to prior-generation models.

Developer Access

Apple shipped two frameworks for developers at WWDC26:

Foundation Models framework: A Swift API giving developers access to the same on-device AFM that powers Apple Intelligence. Also supports cloud models including Claude and Gemini via a Language Model protocol. Free-tier access to AFM 3 Cloud (Private Cloud Compute) for apps with under 2 million App Store downloads.

Core AI: A lower-level framework for bringing third-party on-device models. Zero server dependency, zero token cost, ahead-of-time compilation, fine-grained memory control. Targets developers who want to ship their own model weights on Apple hardware without Cloud access.

What It Means

Apple’s AFM-3 lineup represents a credible in-house model stack — not a Gemini repack, but not independent either. Gemini shaped the weights. Nvidia and Google host the frontier tier. Apple Silicon runs the rest.

The on-device architecture (IFP flash-storage sparse activation) is a genuine contribution to the model efficiency literature. Running 1-4B active parameters from a 20B weight set on consumer hardware, with routing decisions made per-prompt rather than per-token, is a different approach to on-device capability than anything currently shipping from competing labs.

The market implication for Google is mixed. Google gets usage of its Gemini outputs in Apple’s training loop plus server revenue from AFM 3 Cloud Pro. It does not get the headline Siri integration that Bloomberg implied. The Gemini app remains a separate, user-installable product on iOS.

Key Numbers

• 5: AFM-3 models in total (2 on-device, 3 cloud)
• 20B / 1-4B: AFM 3 Core Advanced total / active parameters
• 3B: AFM 3 Core parameter count
• 4: Models running on Apple Silicon
• 1: Model running on Nvidia/Google infrastructure (AFM 3 Cloud Pro)
• 90%: Flash-to-DRAM routing resolved per-prompt, not per-token