Introducing the Third Generation of Apple’s Foundation Models

2026-06-08 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, long

Summary

Apple introduced its third generation of Apple Foundation Models (AFM) on June 8, 2026, a family of five custom-built models developed with Google. These models power Apple Intelligence, deeply integrated into operating systems with privacy at its core, running on-device or via Private Cloud Compute. The family includes two on-device models: AFM 3 Core (3-billion-parameter dense) and AFM 3 Core Advanced (20-billion-parameter sparse, multimodal, activating 1-4 billion parameters). Server-based models include AFM 3 Cloud (workhorse), ADM 3 Cloud (Image) for generation/editing, and AFM 3 Cloud Pro (most capable), which extends Private Cloud Compute to NVIDIA GPUs in Google Cloud. Evaluations show AFM 3 Core improved text preference to 45.6% and AFM 3 Cloud to 64.7% over predecessors. AFM 3 Core Advanced achieved a 4.15 MOS for expressive voices and 44.7% preference for dictation quality.

Key takeaway

For AI Architects and Machine Learning Engineers designing privacy-centric, high-performance AI systems, Apple's third-generation Foundation Models offer a blueprint. You should consider sparse activation architectures like IFP for on-device scalability, enabling large models on consumer hardware by managing memory efficiently. Additionally, explore secure server-side inference via Private Cloud Compute, potentially extending to hybrid cloud environments with partners like Google and NVIDIA, to balance advanced capabilities with stringent data privacy guarantees.

Key insights

Apple's third-gen foundation models use novel sparse architectures and Private Cloud Compute for enhanced on-device and server AI with privacy.

Principles

• Store full models in flash, load experts into DRAM per prompt for on-device scalability.
• Implement architectural refinements like PT-MoE for server-side multimodal reasoning.
• Train with diverse data, excluding user private data, and respect publisher opt-out.

Method

AFM 3 Core Advanced uses Instruction-Following Pruning (IFP) to store the full model in flash memory (NAND) and selectively load a small subset of "experts" into DRAM per prompt, periodically reselecting them during generation.

In practice

• Utilize sparse activation architectures to scale large models beyond traditional DRAM limits on consumer hardware.
• Employ Quantization Aware Training to compress models while maintaining high accuracy for target hardware.

Topics

• Apple Foundation Models
• On-device AI
• Private Cloud Compute
• Sparse Architectures
• Multimodal AI
• Responsible AI

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Everything announced at WWDC26 — Apple Intelligence evolved from isolated features to a privacy-first, ecosystem-wide AI foundation, centered on a rebuilt Siri.
• Liquid AI Releases LFM2.5: A Compact AI Model Family For Real On Device Agents — LFM2.5 offers a compact, multimodal AI model family optimized for efficient on-device and edge deployments.
• On-device AI agents hit a hard memory limit. Apple's new architecture routes around it. — Apple's new on-device AI architecture uses NAND flash for model weights, bypassing DRAM limits for larger local models.
• Apple says its AI is still private, even when it's running on Google's servers — Apple extends its privacy architecture to third-party clouds using hardware-backed confidential computing and strict data transience.
• Building Blocks for Foundation Model Training and Inference on AWS — Foundation model scaling now requires integrated infrastructure for pre-training, post-training, and inference, emphasizing compute, network, and storage.
• Apple bets cheaper AI will woo small developers — Apple is reducing AI infrastructure costs for small developers to foster innovation and attract new talent.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.