Apple wants your iPhone to handle bigger AI tasks on-device; Here's how

Apple has met with startup PrismML, a company known for developing ultra-dense large-language models, to discuss using its technology to run larger artificial intelligence models directly on iPhones. According to The Information report, PrismML has reduced Alibaba’s open-source large language model Qwen 3.6 enough for it to run fully on an iPhone 17 Pro.

Currently, Apple’s AFM 3 Core Advanced model powers iOS 27 features, including the revamped Siri AI voices and improved system-wide dictation on the iPhone 17 Pro and iPhone Air models. It was highlighted that Qwen 3.6 has 27 billion parameters, more than Apple’s on-device AFM 3 Core Advanced model, which has 20 billion parameters. 

Must read: RAM shortage declines global PC shipments down to 4.9%, Apple gains market share

How PrismML’s approach differs

Unlike AFM 3 Core Advanced, PrismML’s version of Qwen 3.6 keeps all of its parameters active at the same time. Whereas Apple’s 20-billion-parameter on-device model uses a sparse architecture, with only 1 billion to 4 billion parameters active at a time.

"One new on-device Apple model has 20 billion parameters but uses a so-called sparse architecture, in which only 1 billion to 4 billion parameters are active at a time," the report said. "In the case of PrismML's on-device model, all 27 billion parameters are active at the same time," it added. Therefore, PrismML's solution offers 7x to 27x higher performance than Apple's current implementation while running on the same iPhone 17 Pro hardware.

Must read: Apple signs $30 billion Broadcom deal in biggest US manufacturing push

Now, Apple is examining the technology because running larger models directly on iPhones could shift more Apple Intelligence features from the company’s Private Cloud Compute servers to on-device processing. This may help lower Apple’s costs and further strengthen user privacy. 

While the technology could benefit Apple in many ways, it may also result in potential backlash, and the company is unlikely to use Chinise AI models for Siri. Therefore, PrismML's solution is likely intended as a proof of concept.