Introduction

On June 9, 2026, Apple stood on a stage at WWDC and announced the third generation of Apple Foundation Models — a family of five AI systems described by the company as the backbone of the next generation of Siri, image editing, and intelligent features across every Apple device. The announcement contained a detail that Apple's marketing team presented quietly but that represents one of the most consequential strategic shifts in the company's recent history: the models were, in Apple's own words, "custom-built in collaboration with Google."

That phrase is doing significant work. The company that built its entire brand on vertical integration — on the principle that hardware, software, and silicon made by Apple, for Apple, produced the best and most private experience — has partnered with the maker of Android to build the AI that powers the new Siri. It has extended its Private Cloud Compute infrastructure, for the first time, to third-party servers. It has run its most powerful cloud model on Nvidia GPUs hosted in Google Cloud. And it has, according to a joint Apple-Google statement from January 12, 2026 that Bloomberg's Mark Gurman estimated to be worth approximately one billion dollars per year, incorporated Google's Gemini technology as the training foundation for its next-generation models.

This is not a minor vendor arrangement. It is a redefinition of what "Apple AI" means.

What the Five Models Are and What Each One Does

The AFM 3 family divides into two distinct tiers: models that run on-device, and models that run in the cloud through Apple's Private Cloud Compute infrastructure. Understanding which model handles which task, and why, is essential to understanding both what Apple built and what it was unable to build on its own.

The first on-device model is AFM 3 Core, the direct successor to Apple's original 3-billion-parameter dense model. It is the everyday workhorse — the model that handles the majority of Apple Intelligence requests on iPhones, iPads, and Macs without ever sending data to a server. Every query that can be resolved locally, is resolved locally, which is the privacy commitment Apple has made and maintained across every generation of the platform.

The second on-device model is the genuinely remarkable one. AFM 3 Core Advanced is a 20-billion-parameter model that runs entirely on-device, on Apple silicon, activating only 1 to 4 billion parameters at a time depending on the complexity of each request. The technique that makes this possible is called Instruction-Following Pruning, developed by Apple's own research team and first described in a paper submitted to academic review in January 2025. Rather than loading the entire model into DRAM — the active working memory of a chip — the full parameter set is stored in flash memory and a lightweight predictor reads each prompt to select which subset of parameters needs to be active for that specific task. Simpler requests activate fewer parameters. More complex requests activate more. The model scales its own footprint in real time, which is how Apple fits a 20-billion-parameter multimodal system onto consumer hardware without draining the battery or requiring hardware that most users do not own. AFM 3 Core Advanced is natively multimodal, handling text, images, and audio within a single model — which is what enables the new expressive voice capabilities and the improved dictation accuracy that Apple reported as a 44.7% human evaluator preference, up from 17.6% in the previous generation.

The three cloud models run on Apple's Private Cloud Compute and handle the tasks that exceed what on-device processing can accomplish efficiently. AFM 3 Cloud is the server-side workhorse, optimised for speed and efficiency across multimodal reasoning tasks. ADM 3 Cloud — where the D stands for diffusion — is the dedicated image model, powering the new Image Playground, advanced photo editing tools, and Genmoji generation. And AFM 3 Cloud Pro is Apple's most capable model, designed for complex reasoning, agentic tool use, and multi-step tasks that require sustained attention and planning across many sequential actions.

The Google Infrastructure Behind the Curtain

The collaboration between Apple and Google is not uniform across all five models. The degree of Google's involvement differs materially between the on-device and server tiers, and understanding that difference is important for evaluating what Apple has traded away in exchange for the capability it has gained.

For the on-device models, AFM 3 Core and AFM 3 Core Advanced, Apple's architecture is its own. The sparse activation design, the Instruction-Following Pruning technique, and the integration with Apple silicon are all Apple Research developments. What Google contributed, according to multiple independent analyses of the technical documentation, is a training refinement signal — Gemini model outputs used to improve Apple's own models through a process known as knowledge distillation, where a more capable teacher model helps a smaller student model improve its responses. Google's Gemini is the teacher. Apple's models are the students. The models that run on your iPhone in production are Apple's own architecture, not Gemini running on Apple silicon.

For AFM 3 Cloud Pro, the picture is different. This is the model that required Apple to make a structural decision it had never made before: to run its AI infrastructure on servers it does not own, in a data centre it does not control, on hardware built by a third party. Apple extended its Private Cloud Compute architecture to Nvidia Blackwell GPUs hosted in Google Cloud, with the security guarantees of that arrangement documented in a joint technical statement from Apple and Google describing a cryptographically verifiable, append-only ledger of all Google Cloud hardware in the Private Cloud Compute fleet. Apple and Google collaborated to build security capabilities that, according to that statement, go beyond a traditional confidential computing deployment — including mitigations for side-channel attacks and supply chain risk documentation that allows external verification of the hardware in the fleet.

The January 12, 2026, joint statement from Apple and Google described the AFM family as "built with Google and its Gemini technology." Apple's June 8 technical post, published on its machine learning research site, used more careful language: "custom-built in collaboration with Google." The practical architecture described across both documents is that Google supplied training infrastructure, frontier model outputs as teacher signals, and GPU infrastructure for the most capable server model, while Apple retained control of the model architecture, the privacy framework, and the on-device stack.

What This Does to Apple's AI Sovereignty

Apple's insistence on building everything independently has been one of the defining characteristics of the company for four decades. It abandoned Intel processors to build its own M-series silicon. It replaced Google Maps with Apple Maps. It built its own operating systems, its own chips, its own retail infrastructure, and its own services ecosystem. The principle underlying all of those decisions is that dependency on a third party creates both competitive vulnerability and quality risk, and that Apple's ability to make products the way it wants to make them requires owning the critical path.

The AFM 3 announcement revises that principle in a specific and significant way. The critical path for frontier AI capability, as Apple's own timeline shows, passed through a dependency it could not resolve internally on the timescale the market required. Apple Intelligence launched in 2024 to widespread criticism that it was underpowered compared to ChatGPT and Gemini. The features Apple had promised — personal context awareness, screen understanding, multi-step Siri actions — were repeatedly delayed through 2025. The company's internal models were not keeping pace with what users expected from the category, and the competitive pressure from Google's on-device AI features, from Microsoft's Copilot integration, and from the rapid improvement of third-party AI products available on iPhone was becoming commercially consequential.

The Google partnership was the resolution of that pressure. At approximately one billion dollars per year according to Bloomberg's reporting, it is an expensive resolution. But it produced AFM 3 Core Advanced, which puts a 20-billion-parameter model on an iPhone — something no competitor has yet accomplished at this scale in an on-device deployment — and it produced AFM 3 Cloud Pro, which gives Apple a server-side model capable of the agentic reasoning and complex multi-step tasks that the new Siri AI needs to actually compete.

How These Models Power the New Siri AI

The features that AFM 3 enables in iOS 26 are the public face of everything described above, and they represent a substantial departure from what Siri has been for most of its fifteen-year existence.

The new Siri AI, built on the AFM 3 family, is designed to understand personal context — the information in a user's messages, calendar, mail, and documents — and use it to answer questions and take actions across apps without requiring the user to navigate between them manually. The features Apple had announced for this capability in 2024 and subsequently delayed are now targeted for delivery on the Gemini-informed foundation. AFM 3 Cloud Pro specifically handles the agentic reasoning required for multi-step tasks: a request to find the confirmation email for a dinner reservation, extract the address, add it to the calendar, and set a departure time reminder involves multiple app interactions, personal data retrieval, and sequential decision-making that requires exactly the kind of sustained reasoning that distinguishes Cloud Pro from the lighter Cloud model.

The on-device models handle the features that must work offline and must not send data to any server: improved dictation, more expressive text-to-speech voices, real-time image understanding for on-device photo tasks, and the everyday conversational interactions that constitute the majority of Siri queries. AFM 3 Core Advanced's multimodal capability on-device is the feature that Apple will use to differentiate the iPhone experience from Android competitors, because running that level of capability without a cloud dependency preserves the privacy story that has been central to Apple's consumer positioning since the introduction of iPhone.

What It Signals About Where Apple's AI Strategy Is Going

The most important question raised by the AFM 3 announcement is not whether the collaboration with Google was the right decision for 2026 — the evidence that it was necessary is visible in Apple's own delay history and the capability gap that the previous generation of models produced. The more important question is what the structure of the partnership reveals about Apple's long-term trajectory.

The January 2026 joint statement described the arrangement as a "multi-year partnership." Apple has not disclosed its expiration date or the conditions under which it could be extended or renegotiated. The terms, described by Gurman as approximately one billion dollars per year, give Apple access to Gemini technology as a training signal and to Google Cloud infrastructure as a server environment. That dependency is real, even if the models running in production are Apple's own. If Google's Gemini capability advances significantly and Apple's internal research cannot keep pace, the asymmetry in the relationship grows. If Apple's own research produces models that make the Gemini training signal less critical, the dependency shrinks. The balance between those two outcomes will determine whether the current arrangement looks, in retrospect, like a strategic bridge or a structural concession.

What is already clear is that Apple's previous model — build everything independently, accept slower progress in exchange for complete control — has been revised. AFM 3 Core Advanced demonstrates that Apple's research team is capable of genuine technical innovation, with the Instruction-Following Pruning architecture representing a meaningful advance in on-device efficiency that competitors have not yet replicated publicly. The decision to partner with Google for training signals and cloud infrastructure is therefore not evidence of a research failure. It is evidence of a strategic recalibration about where the time pressure matters most and where Apple's own capabilities are sufficient to maintain the control it values.

Conclusion

Apple Foundation Models 3 is both more and less than it appears from either direction. It is not, as the most dramatic reading suggests, simply Google's Gemini with an Apple logo. The on-device architecture is genuinely Apple's own, and the sparse activation technique that puts 20 billion parameters on an iPhone is a meaningful technical achievement that required sustained Apple Research investment. But it is also not the fully independent AI platform that Apple's historical strategy would have produced. Gemini's outputs informed the training. Google's infrastructure runs the most capable cloud model. Nvidia's GPUs do the compute. Private Cloud Compute now extends into a third-party data centre for the first time.

The company that built its identity on building everything itself has concluded that, in the specific domain of frontier AI capability, the cost of independence is higher than the cost of partnership. The users who turn on iOS 26 and find that Siri can finally do the things Apple promised in 2024 will likely not care about the provenance of the models underneath. The question that matters for anyone watching Apple's long-term trajectory is whether the company uses the time that partnership has purchased to close the capability gap internally — or whether the billion-dollar-per-year arrangement becomes the permanent architecture of Apple Intelligence.

That answer will not arrive at WWDC. It will arrive in the research papers and product announcements of the next two to three years. For now, the five models are here, the new Siri is coming, and Apple is in a collaboration with Google that neither company's history suggested was likely. That alone is the most significant AI story Apple has told in a decade.