Skip to main content
Apple unveils third-generation foundation model AFM 3 Cloud, showcasing a 36% performance boost in AI processing during a pro

Editorial illustration for Apple unveils third‑gen foundation model, AFM 3 Cloud shows 36% boost

Apple unveils third‑gen foundation model, AFM 3 Cloud...

Apple unveils third‑gen foundation model, AFM 3 Cloud shows 36% boost

2 min read

Apple just rolled out the third generation of its foundation models, a suite it calls AFM 3. The lineup includes five models built in partnership with Google, split between on‑device and server‑based versions. On the phone side, AFM 3 Core is a 3‑billion‑parameter dense model that Apple says offers “a step up in quality.” Its bigger sibling, AFM 3 Core Advanced, pushes the envelope with 20 billion parameters in a sparse architecture that activates only 1 to 4 billion at a time, depending on the request. The model is multimodal, supporting expressive voices and higher‑accuracy dictation, and it runs best on Apple’s most capable silicon.

The cloud side consists of three private‑cloud compute models, designed so user data never leaves the device or is shared with anyone, including Apple. AFM 3 Cloud is billed as the “server‑side workhorse,” optimized for speed and efficiency, while an image‑focused model, ADM 3 Cloud, rounds out the offering. Apple frames the family as the engine behind a new Siri experience and smarter everyday apps, with privacy baked into the architecture from the start.

We also see consistent gains in our single-sided evaluations, which score responses independently along multiple dimensions: AFM 3 Cloud delivers a roughly 36 percent relative improvement in overall response satisfaction and a 21 percent relative improvement in instruction following performance over the 2025 AFM Server model. Further, for image understanding, where the model interprets and reasons over visual inputs, AFM 3 Cloud showed significant improvement over its predecessor from last year, earning preference on 37.8 percent of prompts compared to just 9.6 percent for its 2025 baseline. Finally, AFM 3 Cloud Pro provides an even further improvement over our AFM 3 Cloud, achieving a relative improvement in overall response satisfaction of roughly 10 percent for text and 14 percent for image understanding overall.

Why this matters

Apple’s third‑generation Foundation Models arrive with a claimed 36 percent lift in response satisfaction and a 21 percent rise in instruction following, numbers that catch the eye. Yet we must ask how those percentages translate into real‑world developer workflows. The AFM family, five models spanning on‑device and Private Cloud execution, promises tighter OS integration and a privacy‑first stance, which could ease compliance concerns for apps handling user data.

Collaboration with Google suggests a shared engineering effort, but the exact division of labor remains opaque. For founders, the promise of a ready‑made, high‑performing model may reduce time‑to‑market, though licensing terms and access to the private cloud tier are not detailed. Researchers might appreciate the “single‑sided evaluations” metric, yet without public benchmarks it is unclear whether the gains are reproducible across diverse tasks.

We remain cautiously optimistic: the performance claims are notable, but the practical impact on our projects will depend on transparency, availability, and how Apple balances on‑device constraints with cloud capabilities.

Further Reading