Apple Exploring Ways to Run Much Larger AI Models Directly on iPhones

Apple has held meetings with PrismML astir ways it could usage the startup's exertion to tally overmuch larger AI models straight connected iPhones, according to The Information.


The study said PrismML has managed to shrink down Alibaba's open-source ample connection exemplary Qwen 3.6 to tally wholly connected an iPhone 17 Pro. The exemplary has 27 cardinal parameters, which is larger than Apple's on-device AFM 3 Core Advanced exemplary with 20 cardinal parameters. Apple's exemplary powers iOS 27 enhancements specified arsenic Siri AI's much expressive voices and improved systemwide dictation connected iPhone 17 Pro and iPhone Air models.

Unlike with AFM 3 Core Advanced, each of Qwen 3.6's parameters tin beryllium progressive astatine the aforesaid time.

"One caller on-device Apple exemplary has 20 cardinal parameters but uses a alleged sparse architecture, successful which lone 1 cardinal to 4 cardinal parameters are progressive astatine a time," the study said, successful notation to AFM 3 Core Advanced. "In the lawsuit of PrismML's on-device model, each 27 cardinal parameters are progressive astatine the aforesaid time."

Larger models moving straight connected iPhones would let for much Apple Intelligence features to tally connected instrumentality alternatively of connected Apple's Private Cloud Compute servers, which could trim Apple's costs and further heighten idiosyncratic privacy.
This article, "Apple Exploring Ways to Run Much Larger AI Models Directly connected iPhones" archetypal appeared connected MacRumors.com

Discuss this article successful our forums