Apple iPhones Now Run 20 Billion Parameter AI Models
At the WWDC26 event, Apple officially introduced a groundbreaking advancement in artificial intelligence, confirming that its latest iPhone lineup can now execute models with up to 20 billion parameters. By integrating its new generation of Siri AI with sophisticated architectures trained using Google’s Gemini technology, the company has significantly expanded the on-device processing capabilities of its hardware. This development marks a pivotal shift in mobile computing, allowing high-performance, multimodal artificial intelligence to function directly on user devices without relying solely on cloud-based processing power for complex tasks.
- Apple introduced the AFM 3 Core and AFM 3 Core Advanced models to enhance on-device artificial intelligence performance.
- The new Instruction-Following Pruning technique enables the execution of 20-billion parameter models on hardware with limited RAM.
- Advanced AI capabilities are currently restricted to the iPhone 17 series, specific M4 iPads, and M3-powered Mac devices.
This technological leap transforms the smartphone into a sophisticated edge computing powerhouse capable of managing massive neural networks locally.
Researchers Developed the Instruction-Following Pruning Technique
The primary hurdle for running a 20-billion parameter model on a mobile device is the restricted memory capacity found in smartphones. To circumvent this, Apple researchers engineered a novel method known as Instruction-Following Pruning (IFP). 
Instead of forcing the entire model into the device’s volatile RAM, the system keeps the bulk of the parameters in NAND flash storage. It dynamically fetches only the necessary segments required for a specific query, which typically range between 1 and 4 billion parameters. By shifting from token-based selection to command-based expert selection, the system maintains high efficiency while delivering accurate and context-aware responses.
Hardware Requirements Limit Initial Compatibility
The implementation of this localized processing technology enables features such as high-fidelity expressive voice synthesis and advanced, real-time dictation services. Due to the high demands of these models on unified memory and processing architecture, Apple has restricted availability to its most capable hardware. 
Support is currently limited to the iPhone Air, iPhone 17 Pro, and iPhone 17 Pro Max models.
Beyond the smartphone segment, the company has extended support to iPads equipped with M4 chips that feature at least 12 GB of unified memory, as well as Mac computers running M3 silicon or newer variants. While these on-device capabilities represent the cutting edge of mobile technology, Apple continues to provide cloud-based solutions through its AFM 3 Cloud, ADM 3 Cloud, and AFM 3 Cloud Pro services for more intensive computational needs.
We are curious to hear your perspective on this transition toward more powerful on-device AI; please share your thoughts on whether you prioritize local privacy over cloud-based processing in the comments section below.
Your comment has been submitted,
it will be published after approval.