Apple Introduces Its Third-Generation Foundation Models

Summary

Apple has introduced its third-generation Apple Foundation Models, a family of five models developed with Google to power Apple Intelligence across operating systems. The lineup includes the 3-billion-parameter AFM 3 Core, the 20-billion-parameter sparse AFM 3 Core Advanced for Apple silicon, and three server models running on Private Cloud Compute: AFM 3 Cloud, ADM 3 Cloud for image generation and editing, and AFM 3 Cloud Pro for agentic tool use and complex reasoning. AFM 3 Core Advanced uses Apple’s Instruction-Following Pruning architecture to keep the full model in flash memory, load selected experts into DRAM per prompt, and vary active parameters by task; it activates 1 to 4 billion parameters depending on the request. AFM 3 Cloud builds on Parallel-Track Mixture-of-Experts changes intended to improve training stability, contextual reasoning, and information recall. ADM 3 Cloud supports image creation, editing, Genmoji, multiple aspect ratios and resolutions, and specialized adapters for features such as Spatial Reframing and Image Playground personalization. Apple says training used public, licensed, open-source, study-derived, and synthetic data, but not users’ private personal data or interactions, while publishers can opt out. Pre-training was scaled on newer cloud TPUs, followed by supervised fine-tuning, multi-stage reinforcement learning, hardware optimization, and quantization-aware training. In human evaluations, AFM 3 Core was preferred over its 2025 baseline on 45.6% versus 23.3% of general-text prompts, while AFM 3 Cloud was preferred on 64.7% versus 8.7%. AFM 3 Cloud also showed a roughly 36% relative gain in overall response satisfaction and a 21% gain in instruction following. AFM 3 Core Advanced scored 4.15 for general voice and 4.24 for conversational voice, compared with 3.87 and 3.82 for Apple’s current text-to-speech system, and won overall dictation-quality preferences by 44.7% to 17.6%. Apple says the models run exclusively on-device or through Private Cloud Compute, with AFM 3 Cloud Pro additionally supported by NVIDIA GPUs in Google Cloud under the same privacy guarantees. The models are intended to support a new Siri, photo-editing tools, Image Playground updates, expressive voices, and other Apple Intelligence features; Apple says further evaluations and benchmarks will appear in a technical report later in the summer.