Apple has made a remarkable technological development with the new generation Siri AI that it introduced at the WWDC26 event. The company expanded its on-device capabilities by announcing new AI models trained on Google’s high-end Gemini models.

AFM 3 Core and AFM 3 Core Advanced, among the announced models, are designed to work directly on iPhone and other Apple devices. While AFM 3 Core has 3 billion parameters, the multimodal AFM 3 Core Advanced model stands out with 20 billion parameters.

Innovative Technique: Instruction-Following Pruning

It was a matter of curiosity how a model with 20 billion parameters works on devices with limited RAM capacity, such as the iPhone. To overcome this challenge, Apple researchers developed a new technique called Instruction-Following Pruning (IFP).

Thanks to this technique, the large model with 20 billion parameters is kept in the NAND storage unit instead of being completely loaded into the RAM memory of the device. The system manages the process by transferring only the parts needed for the response to the RAM.

Unlike classical approaches, expert selection is carried out largely on the basis of incoming requests or commands, rather than on the basis of each token. Apple brings the capabilities of the large model to portable devices by loading the specialized part of the 1 to 4 billion parameters required for a specific answer into RAM.

Supported Devices and Features

This new technology allows advanced features such as verbal voices and more advanced dictation to run on the device. Described as Apple’s most powerful on-device model, this system can be used on devices with reasonable hardware needs.

According to official statements, these features are supported on iPhone Air, iPhone 17 Pro and iPhone 17 Pro Max models. In addition, M4 and newer iPad models with at least 12 GB of combined memory and M3 and newer Mac models can also benefit from this technology.

On the cloud side of the company, AFM 3 Cloud, ADM 3 Cloud and AFM 3 Cloud Pro models continue to serve.

What do you think about Apple’s new local process technology?