PrismML Successfully Runs 270 Billion-Parameter Model on iPhone Without Performance Loss, Achieving On-Device AI

PrismML, an American startup, has successfully run a 270 billion-parameter large language model (LLM) on an iPhone, marking the largest AI model executed on a smartphone to date. This achievement could represent a pivotal shift in Apple's on-device AI strategy, reducing reliance on cloud infrastructure while enabling AI processing directly on devices. Using its proprietary compression technology, PrismML executed Alibaba's open-source Qwen 3.6 model on an iPhone 17 Pro, planning to release the model as open-source next week. The technology enables complex tasks like advanced reasoning, autonomous AI agents, and software coding on smartphones without compromising performance, with the model compressed from 54GB to under 4GB. This breakthrough challenges traditional mobile AI architectures, as PrismML's model utilizes all 270 billion parameters simultaneously—a significant departure from sparse MoE structures commonly used in mobile AI. The development offers an alternative to data center-centric AI infrastructure expansion, aligning with Apple's focus on on-device AI to reduce cloud costs and enhance privacy security.