Apple's latest innovation, Apple Intelligence, is powered by two foundational language models: AFM-on-device, a compact yet powerful 3-billion parameter model, and AFM-server, a larger server-based model designed for high-performance tasks. Together, these models bring specialized capabilities—like advanced language processing, coding support for Xcode, and creative visual generation for Messages—into the hands of users and developers.

Apple Intelligence embodies the company's core values, focusing on efficiency, accuracy, and, most importantly, user privacy. Developed with Responsible AI principles at every step, Apple ensures that these models not only enhance user experience but also align with privacy protection and ethical standards.

Here’s the updated list of key facts with the new information:

iOS 18 and Device Compatibility: Exclusive Features

Apple Intelligence, a highly advanced system integrated into iOS 18, is exclusively available on iPhone 15 Pro and iPhone 16, with additional support for MacBook Air and MacBook Pro. The integration leverages a 3-billion parameter model, AFM-on-device, to provide robust language processing capabilities on-device without relying solely on server infrastructure.

Free to Users: No Cost for Enhanced Writing Tools

Apple Intelligence and its Writing Tools, including advanced proofreading, rewriting, and summarizing, are completely free for users, allowing them to access sophisticated AI-driven language features at no cost.

Initial Language Support and Expansion

Starting in October, Apple Intelligence will initially be available in U.S. English, with plans to expand to other languages in future updates. This phased rollout aligns with Apple's commitment to delivering quality and localized AI experiences tailored to user needs.

On-Device Apple Intelligence Model: Built for Privacy and Performance

The AFM-on-device model is specifically designed to operate locally on users’ devices. By running AI models on-device, Apple ensures real-time performance and user privacy, as data remains on the device rather than being processed in the cloud. The AFM-on-device model includes advanced features such as grouped-query attention for efficient memory usage, allowing it to handle demanding tasks while maintaining low latency.

Proofreading Capabilities: Elevate Writing Quality

The Proofread tool goes beyond basic grammar checking. It evaluates word choice and sentence structure, providing real-time suggestions with clear explanations, thereby helping users to enhance their overall writing quality. The AFM model, optimized through knowledge distillation and pruning, makes these real-time capabilities possible while fitting within the constraints of mobile devices【5†source】.

Summarize Tool: Efficient Content Review

The Summarize tool allows users to condense lengthy documents into clear, concise bullet points or key points. This feature utilizes Apple's adapter architecture, allowing on-device models to dynamically adapt to summarization tasks. According to internal evaluations, this summarization capability performs strongly, producing grounded, accurate summaries that follow specific guidelines for different types of content, such as emails and notifications【5†source】.

Enhanced Email Management: Smarter Inbox Features

Apple Intelligence enhances the Mail app by introducing features like Priority Messages, Smart Reply, and quick previews. The AFM models have been fine-tuned using adapter modules specifically for these tasks, allowing users to efficiently handle emails directly on their devices. According to internal benchmarks, AFM-on-device’s adapter for summarization outperformed larger competitor models in human evaluations for comprehensiveness and groundedness【5†source】.

Priority Notifications: Quick and Important Updates

Important notifications are prioritized and displayed in a summarized form on the Lock Screen, allowing users to easily keep up with crucial updates. This feature leverages Apple Intelligence’s capabilities in classification and summarization to highlight what's most relevant without overwhelming the user.

Can Apple Intelligence Be Disabled?

Yes, Apple Intelligence can be turned off in the iOS Settings. However, turning it off might impact other functions like Siri. The relationship between Apple Intelligence and other iOS functionalities is deeply integrated, ensuring a smooth experience for those who leave it enabled.

Privacy by Design: A Foundational Principle

Apple Intelligence adheres to strict privacy guidelines, processing data locally and keeping personal information entirely on the user’s device. By integrating responsible AI principles, Apple prioritizes privacy at every stage of model development, from pre-training through post-training and deployment. No personal data or communications are shared with Apple or third parties, as all processing occurs securely on-device.

Private Cloud Compute (PCC): private AI processing Highlights

1. Advanced AI in the Cloud with Privacy: Private Cloud Compute (PCC) is designed to handle advanced AI tasks in the cloud while keeping user data private. It combines the power of cloud computing with the industry-leading privacy protections Apple is known for.

2. User Data Security: PCC uses custom-built Apple silicon and a specialized operating system to ensure that any personal data processed remains inaccessible to anyone, including Apple staff. Data is only used to fulfill a user's request and is not retained afterward.

3. Stateless Data Processing: PCC processes user requests without storing any personal information. This means data is deleted as soon as a task is completed, ensuring no trace remains within the system.

4. No Privileged Access: There are no "backdoors" or privileged interfaces that could be used to bypass security. This means even Apple engineers can't access personal data during system troubleshooting or maintenance.

5. End-to-End Encryption: All user data sent to PCC is encrypted from the user's device to the cloud, and only authorized PCC nodes can decrypt it. Supporting services like load balancers don’t have access to the encryption keys, maintaining privacy.

6. Verifiable Transparency: Security researchers are able to verify the privacy and security of PCC. Apple will make the software used in PCC publicly available, allowing experts to confirm that the production environment matches security standards.

7. Non-Targetability: PCC is designed so that attackers cannot target specific users. Instead, the system works in a way that makes any attack attempt broad and detectable, making targeted attacks very difficult.

8. Custom Hardware for Privacy: PCC nodes use custom Apple server hardware, similar to the technology used in iPhones. This includes features like Secure Boot and the Secure Enclave, which prevent unauthorized changes and maintain high security standards.