At the recent WWDC26 keynote, Apple showcased its third generation of Apple Foundation Models (AFM), consisting of five distinct models that include both local on-device functionalities and cloud-based applications. A notable development is the inclusion of one model leveraging servers operated by Google using Nvidia chips, reflecting a strategic partnership aimed at enhancing AI capabilities.
Understanding Apple's Foundation Models
Back in 2024, Apple first introduced its foundation models, which featured an initial on-device language model characterized by approximately 3 billion parameters. They also unveiled a larger server model designed to operate within Apple's Private Cloud Compute infrastructure, harnessing Apple silicon to offer cloud-based AI services while ensuring user privacy. This has become even more critical as Apple seeks to address concerns regarding data security in AI applications.
Initially, Private Cloud Compute aimed to safeguard the privacy users expect from on-device processing while expanding into cloud-based functionalities. This ambition required constructing an entirely in-house architecture, which Apple managed to achieve through its dedicated data centers. Apple ensured that the integrity of this architecture could be verified independently by third-party security analysts.
As Apple's AI aspirations have evolved, the company made the decision to collaborate with Google to integrate Gemini into its AI framework, a partnership that has been pivotal in the realization of their latest AI capabilities, especially highlighted at the WWDC26 event.
The New Model Breakdown
The AFM 3 lineup now includes five distinguishable models: AFM 3 Core and AFM 3 Core Advanced, both designed for on-device deployment, and AFM Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro, which operate on cloud servers. The "D" in ADM 3 Cloud (Image) signifies its focus on diffusion technology, which plays a crucial role in generating and editing images.
Of these models, only AFM 3 Cloud Pro operates outside of Apple's silicon infrastructure, utilizing Nvidia GPUs hosted on Google Cloud. This representation marks a milestone as Apple extends its Private Cloud Compute architecture to external frameworks while maintaining stringent security and privacy measures.
Model Features:
- AFM 3 Core: The updated 3-billion-parameter model recognizes a notable quality increase.
- AFM 3 Core Advanced: A powerful on-device model featuring native multimodality, capable of tasks such as extended dictation. It utilizes a 20-billion-parameter architecture activated in portions based on requests, optimizing performance through sparsity.
- AFM Cloud: A robust server-side model aimed at providing speed and performance.
- ADM 3 Cloud (Image): Focused on delivering advanced photo-editing functionalities and new tools like the Image Playground.
- AFM 3 Cloud Pro: Designed for demanding applications requiring complex processing and reasoning capabilities.
Highlights: AFM 3 Core Advanced and AFM 3 Cloud Pro
The standout models here are AFM 3 Core Advanced and AFM 3 Cloud Pro. The former’s integration of 20 billion parameters into an on-device model is impressive, particularly as most public-facing models usually have a significantly lower parameter count. Its efficiency stems from a unique sparse architecture that activates a variable number of parameters (between 1 to 4 billion) based on user queries.
AFM 3 Cloud Pro draws attention as Apple’s leap into external infrastructure while still prioritizing privacy. It harnesses advanced computation techniques from Google’s infrastructure, designed to exceed traditional confidential computing deployments. Highlights of its design include a guarantee against supply chain attacks and a multi-layered security architecture aimed at protecting user data.
Training and Evaluation Methodology
Apple emphasizes that all model training processes incorporated a mixed dataset, drawing from publicly available information, licensed third-party data, and synthetic materials. Importantly, Apple reported that no user data or specific interactions were used during training, offering a level of transparency beneficial for user trust.
To assess the effectiveness of its models, Apple conducted extensive evaluations, where human reviewers measured attributes like instruction adherence, truthfulness, and overall presentation across various applications. Results indicated significant improvements for the third-generation models when compared against their predecessors, showcasing positive evaluations across significant use cases including dictation quality and image understanding.
Conclusion and Future Outlook
The advancements in Apple's third-generation Foundation Models present a notable shift in the AI capabilities of its ecosystem. With enhanced privacy measures and superior on-device and cloud processing, these models reflect Apple’s commitment to delivering advanced, user-centric AI features. It sets a promising precedent for future developments in both personal and enterprise applications, pushing the boundaries of interactive technology while retaining a strong focus on data privacy.
For more detailed insights into the third-generation Apple Foundation Models, check out Apple’s dedicated Machine Learning Research blog.
