Apple in talks with startup that shrinks AI models to run on an iPhone

Apple is reportedly in discussions with a promising Silicon Valley startup, PrismML, which has developed technology designed to compress advanced artificial intelligence models for deployment on iPhones. The startup's CEO shared insights with CNBC, revealing that their innovative approach can shrink a powerful AI model from approximately 54 GB to under 4 GB, enabling it to run effectively on the latest iPhone models. PrismML, a spinout from the California Institute of Technology and backed by Khosla Ventures, recently unveiled compressed versions of Alibaba's open-source Qwen model. CEO Babak Hassibi indicated that Apple, among other tech players, is actively evaluating the speed, energy efficiency, and overall performance of these models on mobile devices. He described the current stage of discussions with Apple as preliminary, emphasizing that progress is being made. This development coincides with Apple's release of the public beta for iOS 27, marking a significant update for users and enhancing Siri's capabilities to better compete with AI assistants from OpenAI and Anthropic. By adopting this technology, Apple aims to keep more data processing on-device, which could mitigate latency issues and potentially lower cloud service costs, all while reinforcing user privacy. Industry analysts believe that compressed models could facilitate more advanced features on iPhones, such as enhanced computational photography and health monitoring tools, which often require sensitive personal data. Carolina Milanesi from Creative Strategies highlighted that the ability to perform more functions locally would greatly benefit users who prioritize privacy. PrismML's compression technique significantly reduces memory requirements by simplifying how data is stored, allowing for faster response times and lower energy consumption when compared to existing models. However, Hassibi noted a slight trade-off in performance, particularly in factual recall when utilizing these models. The startup is offering two versions of its compressed model for free, designed to operate on everyday devices including iPhones and MacBooks. Hassibi revealed that the technology could eventually extend beyond smartphones to robotics and other autonomous systems requiring rapid decision-making. Apple has already implemented several AI functionalities locally, including translation and personal data-related features. Horace Dediu from Asymco suggested that Apple is likely strategizing to retain most common Siri interactions on-device while delegating more complex tasks to cloud processing. Despite the promising nature of PrismML's claims, experts urge caution, noting that the technology's practical performance across various scenarios remains to be validated. Factors such as battery efficiency during multitasking and the model's reliability under extensive use will be crucial for its success. Additionally, this announcement comes amidst ongoing discussions in the tech industry regarding the impact of AI advancements on memory chip demand. While PrismML's approach may lead to reduced requirements for some datacenter operations, analysts caution that the overall need for processing power and memory in AI applications will persist. Ultimately, if successful, this collaboration could transform how AI capabilities are integrated into consumer devices, offering a more efficient and privacy-centric experience for users while bolstering Apple's competitive edge in the evolving AI landscape.