Apple's 5 New AI Models Built from Google Gemini

Apple spent years telling us it was different. Controlled hardware. Sovereign software. Privacy as a product feature, not a policy footnote. The company that refused to open its App Store, fought the FBI over encryption, and built its own chips rather than depend on anyone else just licensed its core AI technology from the maker of Android.

That's not a small thing to paper over with a co-branding agreement.

Key Points

• Apple's new AI runs on Gemini: At WWDC 2026, Apple unveiled five new Apple Foundation Models built on Google's Gemini technology, the result of a multi-year partnership announced in January 2026 reportedly worth around $1 billion per year.
• The branding stays Apple: Apple describes the models as developed "in collaboration with Google." The underlying technology is Gemini. The name on the box is still Apple Foundation Models.
• Five models, two tracks: Two run on-device for everyday tasks, three run in Private Cloud Compute for heavier lifting including image generation, agentic tool use, and complex reasoning.
• Privacy is the load-bearing wall: Apple and Google both state that no user data is stored or shared — not even with Apple. That claim will be tested the moment these features ship widely.
• The new Siri finally arrives: Features Apple announced in 2024 and delayed repeatedly — personal context, screen awareness, multi-step app actions — are now supposed to actually land, on a Gemini foundation.

What Apple Built

The five Apple Foundation Models span a range from a 3-billion-parameter on-device model handling everyday text tasks to AFM 3 Cloud Pro, a reasoning-focused server model optimized for agentic tool use. Training ran on Google's cloud TPUs. Hosting runs through Private Cloud Compute, which Apple continues to position as its privacy differentiator.

The most technically interesting entry is AFM 3 Core Advanced, a 20-billion-parameter multimodal model that would normally be far too large to run on a phone. Apple's solution is Instruction-Following Pruning, an architecture that keeps the full model in flash storage rather than working memory and activates only 1 to 4 billion parameters per prompt depending on the request. In speech synthesis tests, it scored a Mean Opinion Score of 4.15 versus 3.87 for the previous system. In dictation quality, it was preferred in 44.7% of evaluations against 17.6% for the prior model.

At the server end, AFM 3 Cloud Pro is the only model in the family not optimized for Apple Silicon — it runs on NVIDIA GPUs in Google Cloud, an arrangement Apple, Google, and NVIDIA built together with what Apple says are identical privacy protections to the rest of Private Cloud Compute.

The Deal That Rewrote Apple's AI Strategy

Bloomberg's Mark Gurman put the financial scope at roughly $1 billion per year for access to a custom Gemini model with a reported 1.2 trillion parameters. The joint statement from January 2026 said Apple determined that "Google's technology provides the most capable foundation for the Apple Foundation Models" after evaluation. That is a direct acknowledgment that Apple's own models were not good enough, stated in the most neutral corporate language available.

The irony is structural. Google makes Android. Apple makes iOS. These two companies have competed for the smartphone market for nearly two decades. Now the intelligence layer inside every iPhone, iPad, and Mac is running on technology Apple pays Google approximately $1 billion annually to access. The deal that was quietly negotiated while everyone was watching OpenAI turned out to be one of the more consequential partnerships in recent tech history.

What It Means That Apple Is Calling It "Apple Foundation Models"

This is the part worth sitting with. Apple is not saying "powered by Gemini" the way a product says "Intel Inside." It is presenting these as Apple Foundation Models, developed in collaboration with Google, full stop. The Gemini origin is disclosed in the technical details but not in the product narrative.

That choice is deliberate and understandable from a brand perspective. Apple has spent considerable effort building consumer trust around privacy and control. Admitting that your AI runs on your primary competitor's technology is a complicated message to deliver to users who bought into the Apple ecosystem partly on the premise of independence.

Whether that framing holds up is a real question. Independent benchmarks do not yet exist. The models are in beta. The features Apple announced at WWDC 2024 and delayed through 2025 — the ones that were supposed to make Siri genuinely useful — are now promised again, this time on a Gemini foundation. We have been here before.

What Marketers and Growth Teams Should Note

For anyone building campaigns, workflows, or products on top of Apple's ecosystem, a few things follow from this shift. Siri's long-promised capability improvements — screen awareness, personal context, multi-step task execution — are now backed by a model with real capability behind it rather than Apple's prior in-house efforts. If those features ship as described, the on-device AI use case for iOS becomes substantially more interesting for workflow automation and personalization.

The privacy architecture also matters for how brands think about first-party data. Apple's Private Cloud Compute model, if it holds under scrutiny, represents a meaningful constraint on what Apple and Google can access about users. That shapes what behavioral signals will and won't be available for targeting and personalization on Apple devices going forward.

The deeper signal here is about the consolidation happening at the infrastructure layer of AI. The number of companies with the compute, data, and model capability to play at the frontier is very small. Everyone else, including Apple, is now licensing from that group. Understanding who controls that infrastructure — and what the terms of access look like — is increasingly relevant to how you build your AI marketing strategy for the next several years.

If you want help building AI-assisted growth programs that account for where the technology is actually heading, our team at Winsome Marketing is the place to start.