Woofun AI reports that PrismML, co-founded by Caltech mathematician Babak Hassibi, executed the Alibaba open-source Qwen3.6-27B model on an iPhone 17 Pro. The 27-billion-parameter dense model, which utilizes all parameters during inference unlike Mixture-of-Experts architectures, was compressed from approximately 54GB to less than 4GB without significant performance degradation. This optimization enables complex dialogues, reasoning, agent tasks, and coding operations on-device. Apple has engaged with PrismML regarding technology implementation, aiming to advance local AI capabilities as current Siri versions depend on cloud infrastructure.