Apple Exploring Ways to Run Much Larger AI Models Directly on iPhones
MacRumors
Apple's talking to a startup called PrismML about squeezing a 27-billion-parameter AI model onto your iPhone. That's bigger than Apple's own on-device model, and it could mean less reliance on the cloud.
Based on reporting by MacRumors — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Apple doesn't usually go shopping for outside AI help, but The Information reports the company has been meeting with a little-known startup, PrismML, about a technical trick that could reshape how much AI muscle an iPhone can carry on its own. PrismML claims it has taken Alibaba's open-source Qwen 3.6 model, all 27 billion parameters of it, and gotten it running entirely on an iPhone 17 Pro. No cloud round-trip, no Private Cloud Compute server doing the heavy lifting. Just the phone.
That number matters because Apple's own current on-device model, AFM 3 Core Advanced, tops out at 20 billion parameters, and it cheats a little to get there. It uses a sparse architecture, meaning only 1 to 4 billion of those parameters actually fire at once depending on the task. PrismML's approach with Qwen 3.6 keeps all 27 billion active simultaneously, which is a heavier lift computationally but suggests a denser, potentially more capable model can be squeezed into the same silicon and battery budget Apple already ships.
AFM 3 Core Advanced isn't some background footnote either. It's the engine behind the flashier Siri voice work and the improved systemwide dictation Apple is rolling into iOS 27 on the iPhone 17 Pro and iPhone Air. If Apple can license or adapt something like PrismML's compression technique, the obvious payoff is running more of Apple Intelligence locally rather than shipping requests off to its own server fleet. Less server load means lower operating costs for Apple, and keeping data on the device is the kind of privacy claim Apple has built a decade of marketing around.
This fits a pattern that's been surfacing in other recent reporting: Apple also appears to be shopping for AI chip companies, partly to cut its reliance on Nvidia hardware it currently rents through Google Cloud for the toughest workloads. Between chasing chip acquisitions and courting compression startups, the throughline is the same. Apple wants to own more of its AI stack, top to bottom, rather than depend on partners for the parts that actually do the thinking.
None of this is shipping yet. Meetings with a startup are not a signed deal, let alone a shipped feature, and Apple has a long history of quietly evaluating outside tech that never makes it into a keynote. But the direction is telling. Apple's on-device ambitions are scaling up faster than its own models are, and it's apparently fine looking outside its walls, even to an open-source Alibaba model, to get there.
My take — AI-written commentary, not fact-checked reporting
I'll believe the 27-billion-parameter iPhone model when it ships, but the real story here is Apple quietly admitting its own on-device models aren't good enough yet and reaching for an open-source Chinese model to catch up. That's a healthy dose of humility from a company that usually pretends it invented everything in-house, and it's a decent argument for why open weights matter even to the most closed hardware maker on earth.
Read more about this at: MacRumors
