Apple launches third-generation Foundation Models at WWDC 2026: five models from 3B on-device sparse to Pro cloud, co-built with Google

At WWDC 2026 on June 8, Apple introduced the third generation of Apple Foundation Models (AFM 3), a family of five purpose-built AI models co-developed with Google that spans on-device inference and Private Cloud Compute. The flagship on-device model — AFM 3 Core Advanced — is a 20-billion-parameter sparse architecture that activates only 1–4 billion parameters per prompt and stores its full weights in flash memory (NAND), making a 20B model deployable on iPhone-class hardware. The family powers the redesigned Siri and a new generation of on-device Apple Intelligence features.

What's new

The AFM 3 lineup comprises five models across two deployment environments:

On-device (runs locally on Apple silicon):

• AFM 3 Core — 3-billion-parameter dense model for standard Apple Intelligence tasks such as text summarization, writing suggestions, and smart replies.
• AFM 3 Core Advanced — 20-billion-parameter sparse model featuring "a sparse architecture, activating just 1 to 4 billion parameters at a time." "The full model is stored in flash memory (NAND)" rather than requiring all weights resident in active DRAM. Apple describes the model as "natively multimodal, enabling helpful features like expressive voices and higher-accuracy dictation." This is the engine for the most demanding on-device Siri queries.

Server-based (Private Cloud Compute):

• AFM 3 Cloud — General-purpose cloud model optimized for speed, efficiency, and performance.
• ADM 3 Cloud (Image) — Handles image generation and editing; Apple reports "photorealism across diverse subjects and complex lighting conditions."
• AFM 3 Cloud Pro — Most capable model in the family, designed for complex reasoning and agentic tool use, optimized for NVIDIA GPUs.

Performance benchmarks include a Text-to-Speech Mean Opinion Score of 4.15 (versus a 3.87 baseline) and Speech-to-Text quality preferred by users by a margin of 44.7% to 17.6%.

Context

The second-generation AFM shipped in iOS 18 and macOS Sequoia as the original Apple Intelligence rollout — a launch that drew criticism for Siri improvements falling short of marketing promises and key features shipping months late. AFM 3 is Apple's correction: a substantially larger on-device model (20B sparse versus the earlier 3B dense baseline), native multimodal support, and a structured three-tier cloud architecture.

The Google collaboration is a notable structural shift. Apple has historically built its foundation models in-house, with some server-side work done in partnership with Anthropic. AFM 3 marks the first time Apple has publicly disclosed co-building foundation models with Google.

Why it matters

The AFM 3 Core Advanced architecture — 20B total parameters, 1–4B active, stored in NAND flash — is technically significant. Flash-backed inference trades memory bandwidth for capacity: NAND is slower than DRAM, but enables a 20B-parameter model to run on hardware with 8–16 GB of physical RAM. Apple's sparse activation approach reportedly achieves benchmark parity with much larger dense models at a fraction of the compute cost.

With over a billion active Apple devices, AFM 3 represents the largest-scale deployment of sparse on-device language model inference announced to date. For third-party developers building on Apple's Foundation Models framework, the practical effect is a substantially more capable on-device inference substrate with multimodal input support.

The partnership structure — Google on model training, NVIDIA GPUs for cloud inference — signals Apple is moving to a modular AI supply chain rather than end-to-end in-house development, a strategic departure worth watching.