Startup PrismML ran a 27 billion-parameter AI model on an iPhone 17 Pro. It is among the largest models to run on a mobile device. Models that can run on mobile devices typically have active parameters in the single-digit billions.

The Information reported on July 9 that PrismML compressed Alibaba’s open-source large language model Qwen 3.6 and ran it on the iPhone. The company said the model can handle complex dialogue and reasoning, fully autonomous agents and software coding.

Apple has emphasised an on-device AI strategy that covers AI functions on the iPhone while keeping its distance from the race to build AI data centres.

Babak Hassibi (바박 하시비), PrismML CEO, said in an interview, "Imagine a world where in about 3 years, 95 percent of the intelligence you need is used locally on phones, laptops and home appliances, and only the remaining 5 percent goes to the cloud." He added, "Model compression will change the economic structure of AI itself."

PrismML raised $16.25 million in a seed round earlier this year from Khosla Ventures and others.

The company said PrismML sharply reduced the size of Qwen 3.6 using mathematical techniques. It said performance typically declines when models are compressed, but it emphasised that it reduced the model without any performance loss.