Model Releases
Apple Releases AFM 3 Family of Foundation Models
Apple unveils the third generation of Apple Foundation Models (AFM 3), a five-model family spanning on-device MoE architectures to high-performance cloud reasoning.
June 8, 2026
Model ReleaseAFM 3
Introduction
At WWDC26, Apple officially marked a historical milestone in the evolution of edge computing and private AI with the announcement of AFM 3. This third generation of Apple Foundation Models represents a massive leap forward in how intelligence is integrated into the Apple ecosystem, moving beyond simple text prediction to deeply integrated, multimodal, and agentic capabilities.
The AFM 3 family is not a single model but a strategic suite of five distinct models, custom-built in collaboration with Google, designed to scale from local silicon to massive cloud clusters. By optimizing for the entire hardware stack—from the iPhone in your pocket to Private Cloud Compute—Apple is setting a new industry standard for the intersection of high-performance AI and uncompromising user privacy.
- Announced June 8, 2026, at WWDC26
- Collaborative development with Google
- Five-model family architecture
- Unified privacy via Private Cloud Compute
Key Features & Architecture
The AFM 3 lineup is meticulously segmented to balance latency, power consumption, and reasoning depth. The on-device tier is led by AFM 3 Core, a 3-billion-parameter dense model that provides high-quality intelligence for standard tasks without draining battery life. For more demanding local workloads, Apple introduced AFM 3 Core Advanced, a breakthrough in sparse architecture.
AFM 3 Core Advanced utilizes a Mixture-of-Experts (MoE) approach with 20 billion total parameters, but intelligently activates only 1 to 4 billion parameters per request. This allows it to deliver high-tier intelligence while remaining optimized for Apple silicon. Furthermore, Core Advanced is natively multimodal, supporting expressive voice synthesis and high-accuracy dictation, making it a true sensory model for the device.
- AFM 3 Core: 3B parameter dense on-device model
- AFM 3 Core Advanced: 20B parameter MoE (1-4B active) with native multimodality
- AFM 3 Cloud: Optimized server-side workhorse for speed and efficiency
- AFM 3 Cloud Pro: High-capacity model for complex reasoning and agentic tool use
- ADM 3 Cloud (Image): Specialized model for image generation and editing
Performance & Benchmarks
Early benchmarks indicate that AFM 3 is outperforming both its predecessors and key industry competitors in localized intelligence. Specifically, the AFM 3 Core model has demonstrated superior performance in language understanding tasks, outperforming Qwen-2.5-3B and Gemma-3-4B in MMLU and MMMLU scores. This suggests that Apple's vertical integration of software and silicon is yielding higher density of intelligence per parameter.