PrismML says it squeezed a Qwen model from roughly 54 GB to under 4 GB, and that Apple is measuring how fast and how power-hungry such models are on real devices.
Apple is among the companies putting PrismML's model-compression technology through its paces, testing how quickly the compressed models run, how much energy they burn and how they behave on actual devices — according to the Caltech spinout's chief executive.
Babak Hassibi told CNBC of Apple: "They're really evaluating our technology right now." He also drew the boundary himself — the discussions are very early and he does not know where they end up, though, as he put it to CNBC, "things are progressing nicely."
The claim arrives with something concrete attached. This week PrismML made public compressed builds of Qwen, Alibaba's freely available model, which the company says it cut from roughly 54 GB to under 4 GB with all 27 billion parameters retained, small enough to run on an iPhone 15 or newer. Two of the builds are free, aimed at everyday hardware: iPhones, MacBooks and Nvidia-powered PCs.
By PrismML's account, the method is a brutal simplification of how a model stores its internal values, taking each from 16 bits down to one or three possible values. The company says the result uses 10 to 15 times less memory than a conventional version of the same model on the same hardware. The approach came out of Hassibi's academic work; Caltech holds the patents and licenses them exclusively to the startup, which is backed by Khosla Ventures and closed a $16.25 million seed round in March.
PrismML's release landed a day after Apple opened the iOS 27 public beta, the first broad look most iPhone owners have had at the company's long-delayed Siri overhaul. Separately, Apple is trying to keep more personal information and AI processing on the handset — a goal with a hard limit, since the most capable models typically need more memory and processing power than a smartphone can provide.
Some of that already happens: translation, certain summarization tasks and features tied to personal data run locally, while harder requests are handed off to Apple's private cloud or to outside models. Asymco founder Horace Dediu told CNBC that Apple is probably aiming to keep the large majority of ordinary Siri interactions on the phone and send only the most demanding work to the cloud — and that the payoff isn't merely using less memory but squeezing a more capable model into the same physical envelope. "They're trying to figure out how big a model and how clever a model they can fit on the device," Dediu said. Designing the chips and the software together may give Apple an edge in putting such models to work, with tighter control over how AI runs on the device.
None of this is free. Hassibi told CNBC that compressed models typically shed a few percentage points of overall performance, and that factual recall is the first thing to go — reasoning, math and coding hold up longer.
Analysts are not sold yet, cautioning that the performance claims remain unproven outside controlled demonstrations. Counterpoint Research director Tarun Pathak told CNBC the decisive questions are lengthy prompts, battery drain while multitasking and reliability across millions of requests: "The ultimate test will be millions of queries, thousands of device combinations and robust testing at scale." Phil Solis, IDC's client-processor research lead, told CNBC that power consumption may be the biggest open question of all. Until that bar is cleared — and with Hassibi himself calling the Apple discussions very early and their outcome uncertain — this is an evaluation, and nothing more.