Apple is using "Knowledge Distillation" technology to transform Google's large Gemini cloud model into a lightweight component suitable for on-device processing on iPhones.

According to information disclosed on March 25, 2026, Apple's deep agreement with Google allows its engineers to access the full Gemini model in data centers to conduct in-depth inspection and deconstruction of its operational mechanisms. This strategic move marks that Apple has gained greater autonomy in AI, enabling it to directly utilize high-quality computational results and "chain-of-thought" from Gemini as training sources. By simulating the computational process of large models, Apple can train smaller and more secure foundational models of its own.

Apple

The core advantage of this distillation technology lies in significantly reducing hardware computing power requirements and operational costs while allowing the mini model to maintain response speed and accuracy similar to the original Gemini in specific functions. Although the main goal of Apple's Foundation Model team (AFM) is currently to optimize on-device processing experience rather than develop a general large language model directly competing with Gemini, this move undoubtedly strengthens Apple's core strategy of "device-side intelligence."

With subsequent iterations such as iOS 26.4, this distilled on-device model will enhance the responsiveness and privacy protection of native applications like Siri. From an industry perspective, Apple's "deconstruction and reconstruction" of third-party large models not only alleviates performance anxiety related to running high-parameter models on mobile devices, but also signals that the focus of competition in mobile AI is shifting from cloud-based parameter competitions to a low-end attack on on-device execution efficiency.