Apple made a comprehensive update to its artificial intelligence infrastructure by announcing the third generation Foundation Models (AFM) family at the WWDC26 event. The new system, consisting of a total of five different models, includes solutions that work both on the device and on the cloud side. The details announced by the company provide important clues about how Apple Intelligence will work in the future, and show that third-party cloud infrastructure is included in Apple’s artificial intelligence ecosystem for the first time.
Apple’s base model approach was first announced in 2024. At that time, the company was running a language model with approximately 3 billion parameters directly on the device, while offering the larger-scale model through the Private Cloud Compute infrastructure. This system, which runs on Apple Silicon-based servers, aimed to protect users’ privacy expectations despite providing cloud-based artificial intelligence services. In addition, the security features of the system were designed to be verified by independent researchers.
In the intervening period, Apple’s work in the field of artificial intelligence has not progressed at the expected pace, leading the company to different collaborations. Statements made within the scope of WWDC26 revealed that Google Gemini technology plays an important role in Apple’s new generation artificial intelligence efforts.
What does the new Foundation Models family offer?
Apple’s third generation Foundation Models family consists of five models: AFM 3 Core, AFM 3 Core Advanced, AFM 3 Cloud, ADM 3 Cloud (Image) and AFM 3 Cloud Pro. While AFM 3 Core and AFM 3 Core Advanced work directly on the device, the other three models work on the server side. The letter “D” in the name ADM stands for diffusion technology used in image production.
According to the information provided by the company, all models except AFM 3 Cloud Pro have been developed to run on Apple Silicon hardware. AFM 3 Cloud Pro runs on NVIDIA GPUs on the Google Cloud infrastructure. Thus, Apple moved its Private Cloud Compute architecture outside its own data centers for the first time.
Apple states that existing privacy and security standards continue to be maintained during this expansion. The company emphasizes that in the system developed with Google, not only confidential computing technologies are relied upon, but all components from the hardware to the application layer are protected by verifiable security mechanisms.
AFM 3 Core Advanced attracts attention
In the new series, the AFM 3 Core Advanced model stands out in particular. According to Apple’s statement, although this model has 20 billion parameters, it can work directly on the device. This stands out as a remarkable development, considering that most of the local artificial intelligence models developed for mobile devices and personal computers remain at the level of a few billion parameters.
Apple uses sparse architecture to achieve this performance. Thus, the model runs the needed sections instead of keeping all 20 billion parameters active on every request. It is stated that the system only activates 1 to 4 billion parameters at the same time. Although this approach is similar to the Mixture of Experts method in general logic, it is based on a different technique detailed in the “Instruction-Following Pruning for Large Language Models” research published by Apple last year.
AFM 3 Core Advanced also draws attention with its multi-mode structure. The model supports functions such as voice processing, advanced dictation features and more natural voice responses. Apple states that this model has been optimized especially for use in the most powerful Apple Silicon systems.
AFM 3 Cloud Pro running on Google Cloud targets more complex tasks
AFM 3 Cloud Pro is positioned as the most powerful model of the new family. The model, developed for scenarios that require higher processing power, such as complex reasoning processes, advanced tool usage and multi-stage tasks, runs directly on Google Cloud.
According to the technical information shared by Apple’s security team, all Google Cloud hardware used in the system is tracked with cryptographically verifiable recording systems. In addition, multi-layered security mechanisms are implemented to prevent user data from being leaked. Key management, network traffic processing and model running processes are carried out in virtual environments isolated from each other.
On the other hand, Apple states that new models are built on a common foundation. Each model is then customized according to its own usage area. Capabilities such as audio processing, image understanding, long context support and visual production are added in this process.
On the education side, the company; It says that it uses publicly available data, licensed content, open source data sets, data obtained within the scope of private research and synthetic content. Apple specifically emphasizes that no user data or user interactions are used in the training process. In addition, it is stated that internet publishers may prevent their content from being used in model training.
According to the evaluation results shared by Apple, the third generation Foundation Models family exhibits higher performance in areas such as understanding commands, accuracy, presentation quality and image interpretation compared to previous generations. In particular, the AFM 3 Core Advanced model achieved higher preference rates in terms of quality and comprehension accuracy in human evaluations against the company’s existing dictation system. These results show that Apple has created a hybrid structure that continues its on-device artificial intelligence approach with more powerful models while also activating cloud resources when necessary.