Apple Has Full Access to Gemini — and Is Distilling It Into On-Device AI for Siri

Apple's AI strategy has looked opaque from the outside since the company announced its Google partnership for Apple Intelligence last year. A new report clarifies what that partnership actually enables — and the answer is more architecturally interesting than a simple API arrangement.

According to sources cited by The Decoder, Apple has full access to Google's Gemini models and is using that access to perform knowledge distillation: training smaller, on-device models to replicate the reasoning and instruction-following behavior of Gemini, without the parameter count that would make local deployment impractical on iPhone and Mac hardware.

What Knowledge Distillation Enables

Knowledge distillation is a technique where a small "student" model is trained not on raw data but on the output distributions of a large "teacher" model. The student learns to mimic the teacher's behavior — including nuanced reasoning patterns that a model of its size would not independently develop — while operating at a fraction of the computational cost.

The implication for Apple's strategy is significant. Rather than serving every AI request to Gemini cloud endpoints — which creates latency, raises privacy concerns, and requires persistent connectivity — Apple can embed Gemini-caliber capability into a model small enough to run on the Neural Engine of an A-series or M-series chip. The cloud is used to train the on-device model, not to serve it at inference time.

Why This Matters for Apple Intelligence

Apple's privacy positioning has been one of its most consistent differentiators in the AI era. The company has repeatedly emphasized that Apple Intelligence processes most requests on-device rather than in the cloud. But on-device processing is only credible as a value proposition if the on-device model quality is sufficient — and that has been the primary critique of Siri and Apple Intelligence: capable cloud models, but on-device inference that falls short.

If distillation from Gemini is producing on-device models with meaningfully better instruction following and reasoning quality, it closes the gap between Apple's privacy-first delivery mechanism and the capabilities users expect from frontier AI assistants.

The Google Partnership Calculus

The arrangement is mutually beneficial in ways that are easy to overlook. Google gains a prestigious deployment partner and real-world usage signal from hundreds of millions of Apple devices — valuable training data and brand association. Apple gains access to Gemini-quality intelligence for its on-device models without building a frontier model training program in-house — a capability that would require billions of dollars and years of infrastructure buildout.

Neither company has commented publicly on the specifics of the distillation arrangement. Apple has confirmed the Google partnership but characterized it as a "fallback" cloud model option for complex queries rather than a training data source — a description that, if the distillation reporting is accurate, understates the depth of the technical relationship.