Apple is distilling Gemini on-device and turning to Nvidia for cloud privacy

Apple is set to solidify its on-device AI architecture at next month's WWDC, deeply integrating large language models with its custom chips. For basic tasks, Apple is distilling Google's trillion-parameter Gemini model on-device to train lightweight models that reduce network latency; more complex inference will be routed to the cloud.

To protect cloud privacy, Apple has approved the use of Nvidia's confidential computing technology in Google Cloud to encrypt data and models processed by GPUs. Using a third-party cloud marks a shift from Apple's earlier promise that "all non-local AI requests are processed on Private Cloud Compute servers with Apple's own chips," though the company will retain the Private Cloud Compute brand. Meanwhile, Apple is seeking to acquire startups focused on local inference and has identified Liquid AI as a potential target.

On the system level, specific features show that iOS 27 will include a deep Siri redesign. Siri will be integrated into the Dynamic Island and offer a new "Search or Ask" drop-down interface. A standalone Siri chat app will provide complex workflow capabilities similar to Perplexity. Additionally, users can select third-party agents like Google Gemini, OpenAI's ChatGPT, or Anthropic's Claude in the "Extensions" settings.

The camera app will get a dedicated Siri mode to replace the "Visual Intelligence" feature, letting users photograph objects and trigger Google reverse image search, along with a customizable quick-access panel for adjusting exposure, focus, and night mode. The Photos app will introduce generative AI tools like "Extend" and "Reframe." The system also includes grammar checking, Image Playground generative wallpapers, and low-level optimizations for foldable multitasking. Development will focus on bug fixes and performance improvements.