Apple has held meetings with a startup called PrismML to explore whether its compression technology could be used to run substantially larger AI models directly on iPhones, according to a report from The Information published Thursday.
PrismML has already demonstrated the concept in practice. The company compressed Alibaba’s open-source Qwen 3.6 model, which carries 27 billion parameters, down to a size that runs entirely on an iPhone 17 Pro.
That model is larger than the on-device AI Apple currently ships with iOS 27, which uses a 20-billion-parameter model called AFM 3 Core Advanced.
Why the Architecture Difference Matters to Users
The two models are not just different in size. Apple’s AFM 3 Core Advanced uses a sparse design, meaning only 1 billion to 4 billion of its 20 billion parameters are working at any given moment. PrismML’s version of Qwen 3.6 runs all 27 billion parameters simultaneously.
A fully active model of that size can handle more complex requests and hold more context, which typically translates to fewer wrong answers, better follow-through on multi-step tasks, and more natural responses from Siri.
Apple’s current on-device model powers several iOS 27 features on the iPhone 17 Pro and iPhone Air, including Siri’s more expressive voice responses and an improved systemwide dictation system that handles capitalization and punctuation more accurately.
What Changes If Apple Adopts This Approach
Right now, many Apple Intelligence requests that exceed the on-device model’s capacity are routed to Apple’s Private Cloud Compute servers.
Running a larger, more capable model locally would reduce the frequency with which that happens. For users, that means AI features that respond without waiting for a server connection.
For Apple, it means lower cloud infrastructure costs at a time when the company is processing AI requests across hundreds of millions of active devices.
Privacy also factors in. Apple has built its AI strategy around keeping personal data on the device wherever possible, and processing more requests locally rather than sending them to remote servers fits that framework directly. Every query handled on-device never leaves the phone.
PrismML is backed by Khosla Ventures and has not yet announced a formal partnership with Apple. The meetings described in The Information’s report indicate Apple is evaluating the technology, not that a deal has been reached.
Still, the fact that Apple is actively exploring ways to run denser, fully active models on consumer hardware points to where the company wants its on-device AI capabilities to go ahead of future iPhone generations.
