Apple is reportedly using Google Gemini in a much deeper way than most people expected, with the goal of building smaller and more secure on-device AI models for future iPhones. According to a report from The Information, Apple is applying knowledge distillation techniques to study Gemini and turn parts of that large cloud model into lighter systems that can run locally on its own hardware.
The report says the arrangement between Apple and Google goes beyond simply calling Gemini as a back-end service. Apple is said to have access to the model inside its own data center environment, giving its engineers room to examine how the system works and adapt it for Apple’s own product roadmap. The main target appears to be compact AI models that can handle everyday tasks on devices like the iPhone 17 without depending on constant cloud processing.
At the center of the effort is knowledge distillation. In practical terms, Apple’s own external model acts like a student model that learns from Gemini’s outputs and its step-by-step reasoning patterns for narrow tasks. By breaking down that larger system into smaller functional pieces, Apple can train miniature models to respond with similar speed and accuracy in specific scenarios while using far fewer resources.
That approach would give Apple a few obvious advantages. Smaller on-device AI models cost less to run, demand less compute from the device, and reduce the need to send user requests back to remote servers. Just as important for Apple’s brand positioning, local execution also improves privacy because more of the user’s data can stay on the device instead of being processed in the cloud.
The report says this work is being pushed by Apple’s foundation models team, which signals that local processing is becoming a core part of the company’s long-term AI strategy. Rather than trying to clone Gemini as one giant general-purpose assistant, Apple appears to be building a split system: smaller in-house models for frequent local tasks, with larger remote models reserved for more complex requests.
That split may show up in Apple’s next major Siri push as well. The report claims the broader Siri update expected in June will still rely on the full version of Gemini for harder queries that haven’t been distilled, while Apple’s own lightweight models would take over faster day-to-day tasks directly on the device. If that roadmap holds, Apple could end up using Google Gemini as a training foundation while gradually shaping a more private and efficient on-device AI stack of its own.
