Apple is pursuing a fresh approach to AI by exploring technology that allows powerful artificial intelligence models to operate directly on consumer devices like iPhones. This shift aims to reduce reliance on cloud computing, accelerate response times, and enhance data privacy for users.

Central to this effort are ongoing discussions with PrismML, a startup specializing in compressing large AI models so they can run efficiently on mobile hardware. PrismML recently demonstrated the ability to shrink a massive AI model—from about 54 gigabytes to under 4 gigabytes—while preserving most of its processing power. This compression technique significantly lowers memory use, increases speed, and reduces energy consumption on devices such as the iPhone 15 and newer.

PrismML’s CEO Babak Hassibi described the talks with Apple as early but promising. Apple, alongside other technology companies, has tested PrismML’s models to assess their speed, energy efficiency, and overall performance on actual hardware. Despite slight drops in accuracy, particularly in factual recall, the compressed models maintain strong abilities in reasoning, math, and coding. PrismML has made available two compressed versions compatible with iPhones, MacBooks, and certain Nvidia-powered PCs.

This initiative aligns closely with Apple’s broader strategy to upgrade Siri into a more sophisticated and reliable AI assistant. The recent iOS 27 public beta introduced a major overhaul of Siri, emphasizing on-device processing to handle sensitive user data locally. Running AI models directly on devices offers multiple advantages over cloud-based AI:

  • Reduced latency for faster interactions
  • Lower operational costs by minimizing cloud computing needs
  • Support for offline functionality when no internet connection is available
  • Enhanced privacy and security for personal information, including health and medication data

Industry analysts highlight Apple’s unified control of both hardware and software as a competitive advantage. The company’s integration allows it to optimize AI models specifically for its devices, running most routine Siri tasks entirely on-device while reserving only complex requests for cloud processing. This approach aims to strike a balance between performance and privacy.

Experts also underscore the broader implications of on-device AI, pointing to its potential impact on areas like computational photography, video generation, and health applications. The move toward local AI processing reflects a growing trend across the tech sector to safeguard user data while delivering more seamless AI-driven features.