December 28th, 2023 | Updated on November 4th, 2025
Here Are 5 Takeaways:
- Apple’s Innovative Approach to On-Device AI
- RAM Efficiency Over Premium Phones
- Conversational Speeds and Enhanced Performance
- Potential Inclusivity Across iPhone Models
- Industry-wide Efforts to Reduce System Requirements
In anticipation of numerous Android flagship phones integrating on-device generative AI capabilities in 2024, Apple’s approach had remained relatively quiet.
However, a recent discovery by the Financial Times of an Apple research paper sheds light on the company’s strategy for enabling on-device AI processing, particularly focusing on overcoming limitations related to limited RAM.
The research paper outlines a solution for running large language models (LLMs) on devices with restricted RAM capacity.
Apple’s proposed method involves storing “model parameters” and selectively bringing portions of these parameters into the device’s RAM as needed, rather than loading the entire model at once.
This innovative approach allows Apple to efficiently handle models up to twice the size of the available DRAM, leading to a significant boost in inference speed.
The paper highlights a 4-5x increase in speed compared to conventional loading methods on the CPU and an even more substantial 20-25x improvement on the GPU.
This breakthrough in on-device AI processing could potentially pave the way for enhanced performance and capabilities in future Apple iPhones.
A Pathway For Old iPhones To Get On-device AI?
The effectiveness of on-device generative AI is typically enhanced by ample RAM, providing faster read/write speeds compared to the storage solutions in premium phones.
Swift speeds are crucial for on-device AI, ensuring rapid inference times and eliminating prolonged waits for user responses or completed tasks.
Apple’s unique approach addresses this need without requiring excessive RAM, potentially enabling both older and newer iPhones to support on-device generative AI features.
Apple’s strategy offers the prospect of an on-device AI assistant operating at conversational speeds, facilitating quicker image/text generation, and expediting article summarization.
Notably, Apple iPhones, known for having less RAM than many premium Android counterparts, could benefit from this approach.
For instance, the iPhone 11 series boasts only 4GB of RAM, while even the standard iPhone 15 includes 6GB.
It’s worth noting that Apple is not the sole player in the mobile industry working on reducing the footprint of Large Language Models (LLMs).
Recent flagship processors from Qualcomm and MediaTek support INT4 precision, contributing to the downsizing of these models.
As the industry continues to explore innovative approaches, it’s conceivable that on-device AI features could become accessible even on lower-end phones by minimizing system requirements.
