Apple is in early talks with a small Silicon Valley startup that claims it can compress powerful AI models enough to run directly on an iPhone, the company's CEO told CNBC. PrismML, a Caltech spinout backed by Khosla Ventures, this week released compressed versions of Alibaba's open-source Qwen model, shrinking it from roughly 54 GB to under 4 GB. That lets all 27 billion parameters operate on an iPhone 15 or newer. The startup's CEO, Babak Hassibi, said Apple and others are evaluating the models for speed, energy efficiency, and device performance. He called the discussions very early but noted "things are progressing nicely."

The timing matters. Apple just opened the public beta of iOS 27, giving iPhone users their first broad access to a revamped Siri. The company has been trying to catch up with assistants from OpenAI and Anthropic while keeping more processing on the device—a move that could reduce delays, lower cloud costs, and strengthen its privacy pitch. Analysts say smaller models could let Apple move demanding features like computational photography and health tools onto the phone. "The more you can do on device, the better it is," said Carolina Milanesi of Creative Strategies, pointing to sensitive health data users want kept private.

PrismML compresses models by simplifying how internal information is stored, reducing each value from 16 bits to just one or three possible values. The company says this cuts memory use by 10 to 15 times, speeds up responses by six to eight times, and consumes three to six times less energy. Hassibi acknowledged a trade-off: the compressed models lose a few percentage points in performance, with factual recall weakening before reasoning or math skills. The technology could eventually extend beyond phones to robotics and autonomous systems. "It's very important that the intelligence be local and that it can run fast," he said.

Source: CNBC