Apple wants your iPhone to handle bigger AI tasks on-device; Here's how

Apple is in discussions with PrismML, a startup recognized for its cutting-edge large-language models, to potentially enhance the AI capabilities on its iPhones. Reports from The Information indicate that PrismML has successfully optimized Alibaba’s open-source large language model, Qwen 3.6, enabling it to operate fully on the iPhone 17 Pro. Currently, Apple’s AFM 3 Core Advanced model powers key features in iOS 27, including upgraded Siri voices and enhanced dictation on the iPhone 17 Pro and iPhone Air models. Notably, Qwen 3.6 boasts 27 billion parameters, surpassing Apple’s on-device model, which contains 20 billion parameters. What sets PrismML’s approach apart is its ability to keep all parameters of the Qwen 3.6 model active simultaneously. In contrast, Apple’s model employs a sparse architecture, activating only between 1 billion and 4 billion parameters at any given time. This gives PrismML's model an impressive performance boost, potentially delivering 7 to 27 times higher efficiency on the same hardware. Apple is actively considering this technology as it could lead to a significant shift in how AI features are processed, moving from the company’s Private Cloud Compute servers to on-device processing. This transition could not only reduce operational costs for Apple but also enhance user privacy. However, while the advantages are clear, the company must tread carefully to avoid any backlash. It's worth noting that the use of Chinese AI models for Siri remains unlikely, positioning PrismML's technology as a promising proof of concept for Apple’s future AI endeavors.