News

Apple’s Third-Generation Foundation Models Arrive—Built With Google, Running on NVIDIA

Apple's third-gen foundation models arrive: five models, Google collaboration, NVIDIA in Google Cloud, and a 20B on-device model using Instruction-Following Pruning.

Apple just published the architecture behind its next generation of Apple Intelligence. The headline isn’t a single model—it’s a family of five, each engineered for a specific rung of the on-device-to-cloud ladder. What makes this generation different from the two that preceded it isn’t just parameter counts or benchmark scores. It’s the admission, written plainly in Apple’s own research post, that the company built these models in collaboration with Google.

That phrase does heavy lifting. It means Apple’s most capable on-device model, the 20-billion-parameter AFM 3 Core Advanced, owes its existence to training runs on Google’s cloud TPU accelerators. It means the heaviest reasoning model, AFM 3 Cloud Pro, runs on NVIDIA GPUs hosted in Google’s data centers. And it means the privacy architecture Apple spent years designing for its own silicon now has to prove itself on someone else’s hardware.

Five Models, Two Tiers, One Privacy Promise

The lineup splits cleanly. Two models live on your device. Three live in Private Cloud Compute, Apple’s attested, code-audited cloud infrastructure. The on-device pair: AFM 3 Core, a 3-billion-parameter dense model that handles lightweight text, routing, and fast natural-language understanding; and AFM 3 Core Advanced, the 20-billion-parameter sparse model that powers the new Siri, expressive text-to-speech, higher-accuracy dictation, and on-device image understanding.

The cloud trio: AFM 3 Cloud, the server-side workhorse optimized for speed and multimodal reasoning; ADM 3 Cloud, a dedicated image generation and editing model behind Image Playground, Genmoji, and Photos features like Clean Up and Spatial Reframing; and AFM 3 Cloud Pro, the most capable model in the family, built for agentic tool use and complex reasoning.

Four of the five run on Apple silicon. Only AFM 3 Cloud Pro reaches outside—to NVIDIA GPUs in Google Cloud. That single exception is the architectural story of this generation.

How a 20-Billion-Parameter Model Fits on a Phone

AFM 3 Core Advanced is the technical centerpiece. A 20-billion-parameter model normally lives in a data center, not a pocket. Apple’s solution: Instruction-Following Pruning (IFP), a technique its researchers published in 2025. Instead of keeping all weights in DRAM—the fast, scarce working memory—the full model sits in flash storage (NAND). A lightweight predictor reads the user’s prompt and selects a fixed set of “experts” (rows and columns of the feed-forward networks) to load into DRAM for that request. Only 1 to 4 billion parameters activate at once. The rest stay in flash.

This isn’t traditional mixture-of-experts routing, which swaps experts token by token. NAND-to-DRAM bandwidth is too slow for that. IFP makes the routing decision once per prompt, loads the chosen experts, and keeps them resident for the generation. A high percentage of “shared experts” stay always-active; only the “routed experts” swap in and out. The result: 20-billion-parameter quality—Apple’s evals show it matching 9B dense models on math and coding—at the active memory footprint of a 3B model.

The trade-off is hardware gating. AFM 3 Core Advanced only unlocks on Apple’s most capable silicon. Older Apple Intelligence devices get the 3B dense AFM 3 Core instead.

Private Cloud Compute Goes Third-Party

When Apple introduced Private Cloud Compute in 2024, it ran exclusively on Apple silicon in Apple data centers. The 2026 expansion changes that. AFM 3 Cloud Pro runs on NVIDIA GPUs in Google Cloud, wrapped in an extended PCC implementation that layers NVIDIA Confidential Computing, Intel TDX, and Google’s Titan security chip. Apple maintains cryptographic control over which PCC software devices will trust—your iPhone verifies the signature before sending a request. The hardware underneath can belong to Google and NVIDIA; the keys stay with Apple.

Apple’s security research blog describes the arrangement as a sealed vault dropped inside a building it rents from Google, on machinery supplied by NVIDIA and Intel. The landlord owns the building. The vault, its locks, and the rules for who gets in stay Apple’s. More details are available on Apple’s Security Research site.

What the Evaluations Show

Apple’s human evals tell a consistent story. AFM 3 Core improved from 23.3% preference over its 2025 baseline to 45.6% on general text tasks. On image understanding, it won 61% of head-to-head comparisons. AFM 3 Cloud jumped from 8.7% to 64.7% on text, and from 9.6% to 37.8% on image understanding. AFM 3 Cloud Pro adds another 10% relative improvement on text and 14% on image understanding over AFM 3 Cloud, with a 14% edge on math specifically.

On audio, AFM 3 Core Advanced scored 4.15 MOS on general voice and 4.24 on conversational voice—a 0.28 and 0.42 improvement over the production TTS baseline. Dictation preference: 44.7% to 17.6% over the previous system.

What This Means for the Average User

If you have a recent iPhone or Mac, the new Siri, sharper dictation, and expressive voices run locally—no network round-trip, no per-token cost. Image Playground and Genmoji get a quality lift from ADM 3 Cloud. Complex, multi-step requests that exceed on-device capacity route to AFM 3 Cloud Pro in the extended PCC, with the same privacy guarantees Apple has always claimed.

Developers gain image input in the Foundation Models framework—any Swift app can now pass images to the on-device model for captioning, receipt extraction, or UI classification without a cloud call. The framework still costs nothing per inference.

The catch: none of this works in the EU on iPhone or iPad at launch, and mainland China remains excluded. Regulatory uncertainty, not technical limitation, gates the rollout. This mirrors the staggered deployment Apple signaled for iOS 27’s Siri Camera Mode, which also faces regional gating.

FAQ

Are these Google’s Gemini models running on Apple devices?

No. Apple says the AFM family is “custom-built in collaboration with Google”—trained on Google’s TPUs using technology behind Gemini, then refined with Apple’s own data and post-training. The models are Apple’s; the training infrastructure and some model expertise are Google’s.

Does AFM 3 Cloud Pro send my data to Google?

Apple says no. The extended Private Cloud Compute on Google Cloud maintains the same guarantees: stateless computation, no privileged runtime access, cryptographic attestation, and a verifiable transparency log. Apple devices only trust PCC software signed by Apple.

Which devices get AFM 3 Core Advanced?

Only Apple’s most capable silicon. The 20-billion-parameter sparse model requires sufficient unified memory and flash bandwidth to load 1–4 billion active parameters per prompt. Older Apple Intelligence devices fall back to the 3B dense AFM 3 Core.

When will the technical report arrive?

Apple says “later this summer.” The 2025 technical report published in July; expect a similar cadence.

Can third-party developers use AFM 3 Core Advanced?

The Foundation Models framework exposes the on-device model—now with image input—to any Swift app. No API key, no network, no per-token cost. The cloud models (AFM 3 Cloud, ADM 3 Cloud, AFM 3 Cloud Pro) are not directly accessible via the framework. This local-first approach echoes how researchers are running open models on Apple Silicon Macs, as seen with Gemma 4 abliterated variants that fit in 24GB unified memory.