Apple talks to PrismML to run bigger AI on iPhone

Apple is talking to AI startup PrismML to see if it can run bigger AI models directly on iPhones. PrismML, a Caltech spin-off, has a key tech that shrinks model size to about 1/14 of full version and cuts memory use by over 90%.

The tech uses only {-1, +1} for weights with group scaling, not traditional quantization. It keeps accuracy close to FP16 models while boosting inference speed up to 8 times and cutting energy use by 75% to 80%. PrismML already slimmed down Alibaba's open-source Qwen 3.6 (27B parameters) and ran it fully on iPhone 17 Pro.

Apple wants PrismML's quantization skills to improve local AI performance on devices, reducing cloud dependence. This could boost on-device AI on iPhones, enhancing privacy and response speed.