Apple deepens Gemini access
Apple is deepening its partnership with Google to gain fuller access to Gemini models in Google data centers so it can distill knowledge into compact, on‑device models for iPhones. The approach underlines Apple’s strategy to fuse cloud model strengths with optimized local agents for edge AI use cases. (x.com)
The Information’s reporting says Apple has been granted “complete access” to Google’s Gemini instance inside Apple-controlled data centers and can run the model as a teacher to produce smaller, task-specific student models by capturing Gemini’s internal reasoning traces. (macrumors.com) (macrumors.com) Apple’s distilled student models are described as compact enough to run locally and offline on iPhones, enabling features that don’t require round trips to external servers. (appleinsider.com) (appleinsider.com) Bloomberg reporting earlier this year pegged the custom Gemini model Apple licensed at roughly 1.2 trillion parameters and said Apple was preparing to pay about $1 billion per year for the arrangement. (techcrunch.com) (techcrunch.com) Those heavy cloud inferences for Gemini will run on Apple’s Private Cloud Compute infrastructure — Apple has started shipping Houston-built servers and has signalled M5-class Apple silicon is being deployed for its private cloud. (tomshardware.com) (tomshardware.com) The Information’s account adds that Apple can prompt Gemini to output step-by-step reasoning (the “internal computations”) and then use those paired inputs-and-reasoning outputs to train student models that mimic Gemini’s decision process rather than just its final answers. (mactech.com) (daringfireball.net) Apple’s reporting and subsequent coverage note the distilled, on-device variants aren’t broadly rolled out yet, and the broader Siri overhaul (internal codename “Campos”) is slated to be showcased as part of iOS 27 at WWDC the week of June 8–12, 2026. (winbuzzer.com) (winbuzzer.com)
