Apple gains full Gemini access, uses distillation to build lightweight on-device models

Apple has secured broad access rights to Google's Gemini models, according to reporting from The Information. The company now has full access to Gemini within its own data centers and, critically, permission to use knowledge distillation—a technique for extracting capabilities from larger models into smaller ones.

How the distillation approach works

Apple is leveraging Gemini to generate high-quality training data by extracting both answers and reasoning chains from the larger model. This output serves as training data for smaller models that Apple builds internally. The result: lightweight models that deliver identical answers and reasoning paths as Gemini while consuming far less computational resources.

These distilled versions can run directly on Apple devices without requiring cloud connectivity, a key advantage for privacy and latency-sensitive applications.

The strategic play

This approach mirrors tactics allegedly used by Chinese AI companies, but with a critical difference—Apple has paid for legitimate access rights to Gemini's outputs. The arrangement reflects a pragmatic strategy: rather than building reasoning capabilities from scratch, Apple taps Google's foundation models to train its own smaller variants optimized for device-side execution.

According to The Information, Gemini's design around chatbot and enterprise use cases doesn't perfectly align with Apple's Siri integration goals. This mismatch has motivated Apple to continue building its own models in parallel through its Apple Foundation Models team.

Timeline and expectations

Apple is expected to announce new AI features during its Worldwide Developers Conference in June 2026. The distillation work appears designed to power these announcements with practical, on-device capabilities.

The full scope of Apple's Gemini licensing agreement—including pricing, usage restrictions, and exclusivity terms—remains undisclosed.

What this means

Apple is adopting a hybrid approach: leveraging frontier models from established leaders (Google) while investing in proprietary on-device optimization. This reduces Apple's need to develop world-class reasoning capabilities independently while maintaining control over the user-facing models deployed on its devices. The strategy positions Apple to ship differentiated AI features by WWDC without bearing the full R&D cost of training large foundation models from scratch. For Google, the deal provides both revenue and a partnership with a major AI integration partner.