Apple Talks AI Model Compression with PrismML; 27 Billion-Parameter Models to Run on iPhone

Apple is in early-stage discussions with AI startup PrismML to integrate its model compression technology into iPhone, according to PrismML's CEO. The collaboration could enable 27 billion-parameter AI models to run directly on-device, enhancing Siri capabilities and strengthening Apple's privacy positioning.

On July 14, PrismML announced its compression breakthrough: reducing Alibaba's open-source Qwen model from 54GB to under 4GB while preserving the 27 billion-parameter scale for smooth operation on iPhone 15 and newer devices. The technology cuts memory usage by 10–15 times and boosts response speed by 6–8 times, though the CEO acknowledged trade-offs in factual recall versus reasoning and coding abilities.