In this week's Sunday Reboot, Apple's AI model-shrinking talk could have massive benefits, with a chance of also being a billion-dollar deal if it plays its cards right.
Sunday Reboot is a weekly column covering some of the lighter stories within the Apple reality distortion field from the past seven days. All to get the next week underway with a good first step.
Small AI models, big pricing
Back on July 9, a startup in the AI world gained a lot of attention. PrismML used the mythical power of mathematics and science to somehow cut down the size of large language models (LLMs) by a considerable amount.
The process resulted in models like the 54GB Qwen 3.6 being compressed down to an astoundingly tiny 4GB. That's a 27 billion-parameter model being smushed down to a size that would fit on some promotional USB thumb drives from back in the day.
While there was some suspicion that Apple was looking at the startup, it was confirmed by PrismML itself on July 14. In an interview, PrismML CEO Babak Hassibi said that Apple and other companies were evaluating the work of its technology.
Hassibi's public confirmation was an interesting one, especially since Apple has a tendency to NDA everything it does with other companies. While it is unknown if Apple did the same with PrismML, Hassibi seems confident enough in the tech to ignore the possible ire from a potentially massive client.
I say "client" because this sort of important thing in the current AI-centric market warrants acquisition talks. It's something that, if one company buys PrismML and makes the tech exclusive to it, the buyer then has a serious advantage over the rest of the industry.
An acquisition is entirely a possibility for Apple, based on the July 15 report about the reign of John Ternus is believed that Apple is showing signs of changing its acquisition tactics, from spending hundreds of millions on a single company to spending billions.
It's an extremely small change. What's an extra decimal place between friends?
This sort of advantage could well be valued at the billion-dollar level to Apple, if it does acquire the startup. That said, it would also require the startup to be willing to be purchased, too.
Hassibi's interview doesn't really seem like someone willing to sell up. It feels more like he's trying to get more attention from other AI firms to create a bidding war, if the company does want to be sold.
The alternative is that he's trying to secure as many lucrative licensing agreements as possible from AI companies in general. All while remaining independent.
Either route seems like a win for PrismML.
On-device advantage
For Apple, if the technology is sound, it has big benefits, specifically for its aim of increasing the amount of on-device processing.
Apple's current problem is that iPhones don't have massive amounts of memory, nor do most entry-level consumer devices. While some AI-related tasks can be performed with on-device processing, it can get offloaded to a cloud server.
The remote processing can perform much bigger tasks and workloads than a smartphone. Partly because of the higher amount of processing capability, but mostly because the server can have massive amounts of memory to handle the sizable models in the first place.
Siri on the iPhone is better under iOS 27 and macOS 27 Golden Gate.
Apple has managed to use a process of distillation and training to recreate some of Google Gemini's functionality in a smaller iPhone-friendly model. But something like PrismML could be a force multiplier.
Not only could the big models potentially work entirely on an iPhone in the first place, but these distilled models could get even smaller. That frees up more memory for processing the tasks that use these distilled models.
Add in the continuing improvements in AI processing that are coming down the line, and a future iPhone could be an absolute beast.
Better AI on more hardware
There's also the possibility of Apple doing something completely unexpected: expanding Apple Intelligence's reach.
During WWDC, Apple said that the most powerful on-device models would be available only on specific iPhone and Mac models. That includes the iPhone Air and iPhone 17 Pro with 12GB of memory, the M4 iPad also with 12GB of memory, and an M3 Mac with the same memory or better.
Apple says the most powerful on-device AI models will not be on all devices - image credit: Apple
While processing is a factor, memory is certainly going to be another. It's not hard to imagine Apple using the PrismML tech to cut the size of that "most powerful on-device model" to fit into the smaller memory allowances of earlier models.
Extrapolating that further, maybe Apple could expand the availability of Apple Intelligence to older, lower-specification devices.
Sure, there are some tradeoffs, such as much slower processing of tasks. Owners of older hardware would probably be fine with that.
PrismML is a prime opportunity for Apple to take a massive leap when it comes to AI.
If it wants to dominate the field when it comes to on-device processing, it has a chance to take a very major step toward that goal.
Last week's Sunday Reboot talked about the iPhone 17 Pro Max going into a time capsule for 250 years, and pondered if it will emerge in one piece.
