Huge AI Models on iPhones Locally, Apple Evaluates PrismML

Apple has begun meetings with PrismML to understand how to leverage its technology and bring significantly larger artificial intelligence models to iPhones than those currently in use. This is reported by The Information, citing sources familiar with the discussions between the two companies.

PrismML has managed to compress Qwen 3.6, the open-source language model developed by the Chinese giant Alibaba, to run entirely on an iPhone 17 Pro. The model has 27 billion parameters, a figure that surpasses the 20 billion of Apple's current on-device model, the AFM 3 Core Advanced, which powers features of iOS 27 such as more expressive voices for Siri AI and enhanced system-wide dictation on iPhone 17 Pro and iPhone Air.

Fewer Servers, More Privacy: What It Would Mean for Apple Intelligence

The most obvious difference concerns how the parameters are used. The Apple model adopts a sparse architecture, activating only 1 to 4 billion parameters at a time. On the other hand, PrismML's model keeps all 27 billion parameters active simultaneously, an approach requiring much more aggressive compression to fit within the hardware limits of a smartphone.

A larger on-device model would allow Apple to move functions that currently run on Private Cloud Compute servers directly onto iPhones. The advantage would be twofold: lower infrastructure costs and greater user privacy, as data would never leave the device.

PrismML plans to make its open-source model public on Tuesday, July 14, with capabilities that also include software development tasks. This detail strengthens the hypothesis of Apple's concrete interest, as it seeks to close the gap with competitors in the field of large language models.

This is not Apple's first move in this area. The most notable AI acquisition by Apple remains that of Q.ai, acquired for around $2 billion. Regarding Siri AI, the new version of the assistant debuted with iOS 27; Apple opted for collaboration, working together with Google to develop its engine.