Apple’s multi-billion-dollar deal with Google to power Siri with Gemini was announced in January. The hard part is just beginning: shrinking a multi-trillion parameter model to fit inside a phone.
When Apple and Google announced their multi-year AI partnership in January 2026, the headline was simple: Google’s Gemini would replace the underlying intelligence powering Siri and Apple Intelligence features. Reuters and CNBC both reported the deal, with sources describing it as worth over $1 billion annually.
The technical challenge, however, is extraordinary. According to an Ars Technica report from late May, Apple is working on distilling Google’s largest Gemini models — which run across thousands of GPUs in data centers — down to a size that can run on an iPhone’s neural engine, with all the privacy and latency constraints that implies.
What’s new. Ars Technica reports that Apple’s AI and silicon teams are collaborating on aggressive model compression techniques, including quantization, pruning, and knowledge distillation, to bring Gemini-class capabilities to on-device inference. The goal is a version of Siri that can handle complex, multi-step requests entirely on the phone, without sending data to the cloud.
A cloud component is probably inevitable — Apple’s current Apple Intelligence architecture already uses a “Private Cloud Compute” layer for requests that exceed on-device capability. But the company is pushing hard to maximize what runs locally, motivated by both privacy commitments and the user experience benefits of low-latency responses.
The timeline, according to earlier reports, targets a significantly upgraded Siri with iOS 27, expected in 2027, with incremental improvements rolling out in iOS 26 updates later this year.
The key angle. The Apple-Google Gemini deal is strategically unusual. Apple has spent years building its own silicon (the A-series and M-series chips) and its own machine learning frameworks (Core ML, ANE). Turning to a competitor for the foundational AI model is a concession that Apple’s in-house AI efforts have not kept pace with the frontier.
But it is also a pragmatic move. Google’s Gemini family spans multiple model sizes, giving Apple options: the full model for cloud inference, a distilled version for on-device, and potentially custom-sized variants for different Apple devices. The Apple Neural Engine, now in its third generation, is specifically designed for this kind of on-device AI workload.
Context / What’s next. The first public result of the partnership appeared in iOS 26.4, released in April 2026, which included Gemini-powered Siri features for a limited set of tasks. The full rollout, however, depends on Apple’s ability to shrink the model without losing too much capability. If successful, it would be the most capable on-device AI assistant ever shipped on a smartphone — potentially leapfrogging both Google’s own Pixel-exclusive features and Samsung’s Galaxy AI.
If Apple cannot get the distillation right, the fallback is a cloud-dependent Siri that works well on fast connections but fails in the scenarios where an assistant is most useful: offline, in low-signal areas, or when privacy demands local processing.
The big picture. The Apple-Google Gemini deal represents a rare admission from Apple that it cannot do everything in-house. But it also sets up a fascinating dynamic: Google, which sells the Pixel phone as a showcase for its AI, is now powering the AI features of its biggest hardware competitor. The quality of on-device Gemini on an iPhone will be a direct test of whether Google’s AI advantage can survive outside its own ecosystem.
Sources: Ars Technica (May 28, 2026); Reuters (January 12, 2026); CNBC (January 12, 2026)