Apple Intelligence, The Future of On-Device AI with iOS 18

Apple has revealed that its new on-device AI model, dubbed Apple Intelligence, will initially require 4GB of device storage, roughly equivalent to the size of an HD movie. However, the company has hinted that this storage requirement will increase as more features are rolled out in future updates.

This information was shared in Apple’s “Introduction to Apple Intelligence on iPhone,” ahead of the official public release of iOS 18.1 next month. This release will mark the formal introduction of Apple Intelligence to both iOS 18.1 and macOS Sequoia 15.1.

 

Hardware Requirements

 

To run Apple Intelligence, users will need an iPhone 16 series (iPhone 16, 16 Plus, 16 Pro, and 16 Pro Max) or an iPhone 15 Pro/Pro Max. Additionally, iPad and Mac devices with Apple Silicon M1 or later will also support the AI service.

 

The iPhone 15 Pro and the iPhone 16 models all come with a minimum of 128GB of storage, providing sufficient space for the AI service and other data. However, early testers of beta versions have reported that model sizes vary, with some models taking up 5GB and others significantly less. Apple has also confirmed that storage requirements will increase as additional features are introduced.

 

Apple Intelligence Beta and Rollout

 

Apple Intelligence won’t be fully available at the initial release of iOS 18.1. Instead, it will launch as a beta service, and users will need to join a waitlist for activation. Once activated, the necessary neural network models will be downloaded onto the device. Apple has not confirmed whether users will have the option to remove these models, but deactivating the service is likely an option.

 

The Apple Foundation Model

 

At the core of Apple Intelligence is the Apple Foundation Model (AFM), a generative AI model with roughly 3 billion parameters, designed to run efficiently on mobile devices. AFM is complemented by a server-based language model hosted within Apple’s Private Cloud Compute infrastructure. In addition to the text-based AI capabilities, Apple has developed a diffusion model for adding graphics in Messages and a specialized model for coding in Xcode.

 

Apple has worked to reduce the size of AFM by quantizing its parameters to 4 bits or less, with some layers receiving as little as 2-bit quantization. According to Apple’s research, this approach allows the model to maintain quality while using just 3.7 bits per weight (bpw) in production, significantly reducing its size.

 

For comparison, a 3 billion-parameter model would typically require 6GB of storage at 16-bit quantization, 3GB at 8 bits, and about 1.5GB at 4-bits. This optimization makes the model more feasible for on-device use without significant performance trade-offs.

 

Modular AI with “Adapters”

 

To add flexibility to its AI system, Apple employs “adapters,” which are small sets of additional model weights that can be loaded or unloaded dynamically to offer specific features. These adapters, which occupy only tens of megabytes, enable a variety of capabilities such as text summarization, email replies, and proofreading.

 

Initial Capabilities of Apple Intelligence

 

At launch, Apple Intelligence will offer a range of features, including:

 

1. Writing tools
2. Photo cleanup
3. Creation of Memory movies in Photos
4. Natural language search in Photos
5. Notification summaries
6. Focus features like reducing interruptions and intelligent breakthrough
7. Priority mail in Messages and Mail
8. Smart replies in Messages and Mail
9. Siri improvements (including enhanced product knowledge, better request handling, a new interface, and a more natural voice)

 

Early Reactions and Future Prospects

 

Feedback from beta testers has been mixed, with some describing the service as underwhelming. However, as the AI service becomes more widely available and additional features are introduced, Apple Intelligence may find more compelling use cases.