What are Apple Foundation Models?
Apple introduced five new Foundation Models artificial intelligence models at WWDC26. Details of the on-device and cloud-based hybrid artificial intelligence strategy are in our news.
Apple expanded its artificial intelligence strategy within the scope of the WWDC26 event and officially announced the third generation Apple Foundation Models (AFM) series. With this comprehensive update, the company introduced a total of five different artificial intelligence models to the technology world, some of which run on local devices and some in the cloud infrastructure. Apple aims to push the performance limits with new cloud-based models powered by Nvidia chips, especially by moving its privacy-focused Private Cloud Compute structure to Google’s information centers. These models, which enable users to benefit from both on-device speed and the complex processing power offered by the cloud, begin a new era in Apple’s artificial intelligence ecosystem.
- Apple has introduced five new AI models that combine on-device and cloud-based processes.
- The company has expanded its Private Cloud Compute infrastructure to use NVIDIA GPUs on Google Cloud.
- New models were developed without using user information for training, and confidentiality standards were maintained.
- AFM 3 Core Advanced can manage 20 billion parameters efficiently thanks to its sparse architecture.
What are Apple Foundation Models? How Does It Work?
New generation models; It is classified as AFM 3 Core, AFM 3 Code Advanced, AFM 3 Cloud, ADM 3 Cloud and AFM 3 Cloud Pro. While the first two of these models are optimized to run directly on Apple devices, the others are server-based. In particular, AFM 3 Core Advanced draws attention with its 20 billion parameter structure.
Using sparse architecture, Apple enabled a model of 20 billion parameters to run on the device.
Security Standards Are Maintained
Apple emphasizes that it does not compromise on security when extending its Private Cloud Compute architecture to third-party infrastructures. Models running on Google servers are subject to a transparent and auditable security protocol, as is the case with Apple’s own hardware. This collaboration is built on a layered defense mechanism that supports end-to-end confidentiality guarantee in artificial intelligence processes.
Education Processes Are Conducted with Transparency
Publicly available information, licensed content and synthetic data were used in the training phase of the new models. Apple underlines that no user data or interactions are involved in training these models.
Performance tests prove that the new generation models show significant improvement in both visual and language skills compared to previous versions.
Performance evaluations of the new models show a clear superiority compared to previous generations in every category.
What do you think of Apple’s strategy of combining on-device intelligence with cloud-based AI models? Don’t forget to share your intentions about this major transformation in the Apple ecosystem with us in the comments section.
