Apple is reportedly in discussions with a startup that has developed technology capable of significantly shrinking large artificial intelligence models, enabling them to operate directly on devices like the iPhone. This breakthrough could enhance features such as Siri by increasing privacy and improving performance through greater reliance on local processing rather than cloud servers.
The Technology Behind On-Device AI
PrismML, a spinout from the California Institute of Technology backed by Khosla Ventures, revealed its ability to compress powerful AI models. The company demonstrated reducing Alibaba‘s open-source Qwen model from an approximate size of 54 GB down to less than 4 GB. This compression allows all 27 billion parameters of the model to function on devices running iPhone 15 or newer.
The technology functions by simplifying how the AI models store their internal data, reducing each numerical value from 16 bits to just one or three possible values. The startup claims this process results in compressed models that require between ten and fifteen times less memory, generate responses six to eight times faster, and consume three to six times less energy compared to conventional versions running on current hardware.
While the CEO of PrismML, Babak Hassibi, noted that the discussions with Apple are “very early” and progressing well, he cautioned that there is a trade-off. The compressed models can lose some overall performance, particularly showing a weakening in factual recall before skills like reasoning, mathematics, and coding.
Implications for Apple’s AI Strategy
The development arrives as Apple rolled out the public beta of iOS 27, an update that includes a major overhaul of Siri. Traditionally, the most advanced AI models have demanded too much processing power and memory to run effectively on smartphones. By enabling more complex AI functions locally, Apple can address this core constraint.
Analysts suggest that keeping common tasks—such as computational photography, video generation, and health-related tools involving sensitive personal data—on the device offers distinct advantages. As Creative Strategies’ president and principal analyst, Carolina Milanesi, observed, “The more you can do on device, the better it is,” especially when handling private user data.
For Apple, running AI locally would not only reduce latency associated with sending data to remote servers but could also lower cloud computing costs and bolster the company’s commitment to user privacy. Horace Dediu, founder of Asymco, stated that “They’re trying to figure out how big a model and how clever a model they can fit on the device.”
Industry Analysis and Market Outlook
While PrismML has released two compressed versions of its model for free use across various devices, including iPhones and MacBooks, industry experts recommend caution. Tarun Pathak, research director at Counterpoint Research, stressed that real-world testing is essential, noting that performance over lengthy prompts, battery consumption during multitasking, and reliability across millions of queries will be critical tests.
The technology also fuels a broader debate regarding the future demand for memory chips. Morgan Stanley estimates that Apple’s average dynamic random access memory cost per bit could rise approximately 190% year-over-year in fiscal 2027, while NAND costs are expected to increase by about 180%. However, some analysts believe these efficiency gains will not eliminate the need for compute power.
“It’s not that you’re not going to need the chip,” stated Gil Luria, an analyst at D.A. Davidson. “You’re still going to need the GPU, and you’re still going to need the memory.” He added that running AI on individual devices could actually be less efficient than using shared datacenter infrastructure.
Ultimately, experts agree that a hybrid approach is likely most effective: complex tasks can be handled by the cloud, while sensitive or latency-critical functions are executed locally. This combination aims to deliver an experience that is more complete, private, and highly efficient.
