Your phone's assistant wants to be smarter, but there is a problem. The most powerful AI models are too big to fit on a device the size of an iPhone. They live in the cloud, which means every time you ask Siri something, your request leaves your phone and travels to a data center. That takes time, uses energy, and raises questions about privacy.
Apple has been trying to fix that. The company wants Siri to be faster and keep your data on your device. But the models that give AI its brains are enormous. Think of them as huge files that need a lot of storage and a lot of memory to run.
That is where PrismML comes in. The startup says it can squeeze those massive models down until they fit onto a phone. Babak Hassibi, PrismML's CEO, told CNBC, "They're really evaluating our technology right now."
How small can they go? The startup took a model from Alibaba called Qwen that had 27 billion parameters, compressed it from 54 gigabytes to less than 4 gigabytes. That is roughly the difference between a full-length movie and a single song.
The technology emerged from the California Institute of Technology, which holds the underlying patents. The startup is backed by Khosla Ventures and raised $16.25 million in seed funding in March.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
Apple is not just looking for better AI. It is also trying to control costs.
For fiscal 2027, Morgan Stanley projects that Apple's per-bit DRAM cost might increase by about 190% compared to the previous year, with NAND costs up about 180%. Memory is a big part of what makes a phone expensive. The analysts anticipate that Apple will increase the base price of equivalent iPhone 18 devices.
Now add the AI problem. Running smart models on a phone requires more memory, not less. Still, doing more on the device means Apple can lean less on cloud services. That saves money on servers and data transfers. Horace Dediu, founder of Asymco, described the challenge this way: "They're trying to figure out how big a model and how clever a model they can fit on the device."
PrismML's technology is not a done deal yet. Apple and other companies are now running their own tests. They are checking whether the compressed models actually deliver the speed, energy savings, and accuracy the startup claims.
Tarun Pathak, who serves as research director at Counterpoint Research, offered a blunt assessment: "The ultimate test will be millions of queries, thousands of device combinations and robust testing at scale." He added that a blend of cloud and on-device AI can provide a more complete, efficient and privacy-centric experience.
PrismML plans to next compress Google's open-source Gemma model, and eventually tackle much larger models that today generally require datacenter hardware. That could open the door for phones that do not need to phone home at all for smart answers.
The bottom line: If the tests pass, Apple could ship an iPhone that runs serious AI without the cloud. That means faster responses, better privacy, and fewer server costs. But those memory cost increases will still hit, so phone prices could rise either way. Carolina Milanesi, president and principal analyst for Creative Strategies, encapsulated the reasoning: "The more you can do on device, the better it is."
For investors, this is worth watching. If Apple pulls it off, the next iPhone could feel noticeably smarter. If not, the wait continues. Either way, the pressure to fit more brain into a pocket-sized computer is only getting stronger.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
