During the WWDC26 keynote, Apple unveiled its latest generation of Apple Foundation Models (AFM), now in its third iteration, featuring a total of five distinct models. This lineup includes both local and cloud-hosted models, with one model operating on Google's servers leveraging Nvidia hardware. Here's a closer look at what each model offers.
To provide some context, Apple first introduced its foundation models back in 2024, which featured an on-device language model with approximately 3 billion parameters, alongside a more robust server-based model that utilized Private Cloud Compute technology running on Apple silicon. The ambition behind Private Cloud Compute was to provide cloud AI capabilities while ensuring the privacy users expect from on-device processing, all managed in Apple data centers.
However, as Apple faced challenges advancing its AI initiatives, it formed a partnership with Google to integrate Gemini into its AI framework, a development highlighted in the recent WWDC26 announcement.
The third generation of AFMs consists of five models: AFM 3 Core and AFM 3 Core Advanced as on-device models, alongside AFM Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro as server-side options. Notably, the "D" in ADM 3 Cloud (Image) represents the diffusion technology utilized in these models.
Apart from AFM 3 Cloud Pro, which operates on NVIDIA GPUs hosted in Google Cloud, all other models are optimized for devices powered by Apple silicon. This collaboration marked the first time Apple has expanded its Private Cloud Compute architecture to include external infrastructure, while still maintaining its stringent security and privacy protocols.
Here’s a summary of each model:
- **AFM 3 Core**: It's the latest iteration of the original 3-billion-parameter model, delivering improved quality. - **AFM 3 Core Advanced**: The most potent on-device model, featuring 20 billion parameters and optimized for higher accuracy and expressive functions, utilizing a sparse architecture that activates between 1 to 4 billion parameters based on the request. - **AFM 3 Cloud**: Serving as the foundational server model, designed for speed and effectiveness. - **ADM 3 Cloud (Image)**: Geared towards image generation and sophisticated editing functionalities. - **AFM 3 Cloud Pro**: The most advanced server-based model, suitable for complex reasoning and demanding tasks.
The standout models are AFM 3 Core Advanced and AFM 3 Cloud Pro. AFM 3 Core Advanced, with its 20 billion parameters, significantly enhances on-device processing capabilities, using a strategy that activates a subset of parameters to optimize performance. This approach, similar to the Mixture of Experts model, employs a concept Apple developed in its research a year prior, aimed at streamlining how large-language models function.
As for AFM 3 Cloud Pro, it relies on third-party infrastructure, with extensive security measures detailed in Apple’s latest Security blog. The collaboration between Apple and Google aims to fortify protective measures against sophisticated attacks, ensuring that the entire operational stack from firmware to application code is secured and verifiable.
In its Machine Learning Research blog, Apple emphasizes that all five models emerged from a common foundational framework tailored for specific use cases, incorporating multimodal capabilities such as audio and image processing as well as advanced reasoning and visual generation. Importantly, the training data consisted of diverse sources while deliberately excluding user data, allowing web publishers an option to opt-out of involving their content in model training.
Apple has conducted meticulous evaluations of its third-generation models, analyzing performance across several factors like instruction adherence and truthfulness. The evaluations included comparisons against previous models, yielding promising results across various categories.
For more detailed insights into Apple’s third-generation Foundation Models, further information can be accessed through their dedicated resources.
