Apple's New Siri Will Be Powered by Google Gemini (www.macrumors.com)

Apple has struck a major AI licensing deal: its next-generation Siri will be powered by Google’s Gemini, a 1.2 trillion-parameter model, under a reported ~$1 billion-per-year arrangement. Apple will use Gemini for heavy-lift tasks such as summarization and multi-step planning/execution while retaining its own models for other features; the Google-hosted model will run on Apple’s Private Cloud Compute so Google reportedly won’t see Apple user data. The move follows Apple’s delay of a Siri overhaul (now expected in iOS 26.4, spring 2026) and comes after Apple evaluated OpenAI and Anthropic options, rejecting Anthropic for cost reasons. Apple is also developing a ~1 trillion-parameter in-house model, aiming to transition away from Google once its LLMs reach parity. For the AI/ML community this highlights two trends: commercial-scale model licensing and pragmatic use of sparse, mixture-of-experts (MoE) architectures to deliver massive capacity without prohibitive inference cost—Gemini’s trillion+ parameters are “sparse,” meaning only a subset activate per query. It underscores that parameter count isn’t the sole signal of capability (architecture and training matter), and it signals continued vertical stratification: companies will mix third-party giant models with proprietary components to balance capability, privacy, and cost while racing to build competitive in-house LLMs.