Apple AFM 3: The Triumph of On-Device AI… and What It Gave Up

On June 8, 2026, Apple unveiled the third generation of its Apple Foundation Models. In them, the company deploys one of the most ingenious on-device architectures on the market. On that very same day, it also announced that Europeans would not get to use it on the device where they would reach for it most, and that the whole family was co-built with Google. Three pieces of news, one story: that of a company still brilliantly in command of your devices’ memory, but which has given up its grip on something else.

Let us take them in order, starting with what genuinely deserves admiration.

The Wall Nobody Shows You at a Keynote

A large language model is a memory problem before it is a compute problem. A 7-billion-parameter model in half precision already demands more than 14 GB just to load its weights, which is more RAM than most phones have. The standard approach is to load everything into DRAM, the fast working memory, and that mechanically caps the size of the model you can run locally.

A smartphone, however, carries two kinds of memory with opposite properties. DRAM is fast but scarce and expensive. Flash memory (NAND) is abundant and cheap, but its bandwidth to DRAM is slow. The entire art of on-device AI plays out in that asymmetry, and that is precisely where Apple planted its flag.

The Stroke of Genius: Routing by Prompt, Not by Token

The flagship on-device model, AFM 3 Core Advanced, carries 20 billion parameters but activates only 1 to 4 billion at a time. So far, nothing new: this is the principle of a sparse architecture, a Mixture-of-Experts in which only a subset of the model works on any given request. The real find lies elsewhere.

Rather than forcing the entire model into DRAM, Apple stores it in flash and loads only the experts it needs. The catch is that a conventional MoE selects its experts token by token. On a phone, that is impossible: the NAND-to-DRAM bandwidth cannot keep up, and you would spend all your time waiting on transfers. Apple’s solution, then, is to make the routing decision once per prompt rather than continuously. A lightweight dense block picks a fixed set of experts at the start of processing, then periodically reselects them during generation. To cut data movement even further, the model leans on a high proportion of shared experts that are always active, supplemented by routed experts loaded into DRAM only when needed.

The side benefit is subtle: elasticity at inference time. Instead of a single model for every task or a swarm of small ones, AFM 3 Core Advanced tunes its number of active parameters to the difficulty of the request. Weights load incrementally, the model scales past the limits of DRAM, and latency stays contained. It is elegant, and it pays off for both battery life and heat.

I will say it without reservation: this is very fine engineering. But it is no free lunch, and above all it is no spontaneous breakthrough.

“Novel”? The Lineage the Marketing Erases

Apple calls this architecture new. The word deserves an asterisk. Back in December 2023, eight Apple researchers published “LLM in a flash: Efficient Large Language Model Inference with Limited Memory” (arXiv 2312.11514, later presented at ICLR). The paper already laid out the whole mechanism: store the weights in flash, bring into DRAM only what serves the task, by way of two techniques (windowing, which reuses already-activated neurons to reduce transfers, and row-column bundling, which reads flash in large contiguous chunks to play to its sequential strengths). The result claimed at the time was running models up to twice the size of available DRAM.

Graft Instruction-Following Pruning on top, another in-house technique that dynamically selects the relevant sub-networks, and you get almost exactly what Apple now presents as a breakthrough. This is not a lie; it is a narrative. AFM 3 Core Advanced is the industrial culmination of a two-and-a-half-year research track, not a rabbit pulled from a hat. The distinction matters, because it separates those who actually follow the R&D from those who swallow the keynote storytelling whole. On the same memory-compression logic, incidentally, there are rival approaches I have already dissected, such as TurboQuant, with which Google cuts LLM memory by a factor of six.

The Trade-Off You Are Not Billed For

Routing by prompt rather than by token is not a superiority; it is a compromise. An MoE that routes at every token maximizes adaptivity: it can summon the ideal expert for each generated word. By freezing the set of experts for the duration of a prompt (with a few reselections along the way), Apple gives up that fine grain. For a conversational assistant, the quality impact is probably marginal. On reasoning that shifts sharply within a single response, the model most likely leaves some performance on the table. In other words, this is a latency-for-quality trade-off dressed up as a feat of memory management. That said, let us be fair: on a phone, this trade-off is probably the right one. Apple is not optimizing for the top of a synthetic benchmark, but for the experience as the end user perceives it, which is to say for battery life, thermal dissipation, and felt latency. Sacrificing a few points of marginal quality for an assistant that answers fast, does not melt the device, and does not drain it in an hour is a clear-eyed engineering decision, not laziness. And this is exactly where the system’s elegance shows: when a prompt exceeds what the on-device model handles comfortably, the orchestration shifts quietly to the server models, AFM 3 Cloud and then AFM 3 Cloud Pro, on Private Cloud Compute. The user sees nothing, the phone stays cool, and the heavy request goes off to be processed elsewhere. Seamless, yes. But hold on to that “elsewhere,” because it is exactly the point at which local mastery hands off to an infrastructure Apple no longer entirely owns, as we are about to see.

The same logic governs compression. Apple turned to Quantization Aware Training to shrink the models substantially while preserving accuracy. This is the state of the art, it works, and it is also an admission: without these crutches, the model would not fit inside an iPhone’s thermal and memory envelope. Here, technical virtuosity measures the size of the constraint; it does not abolish it.

The Server Side, or Sovereignty on a Sliding Scale

This is where the story turns. The AFM 3 family comprises five models. Two run on the device. Three run on Private Cloud Compute servers: AFM 3 Cloud (the workhorse), ADM 3 Cloud for image generation and editing, and AFM 3 Cloud Pro, the most capable, devoted to agentic tool use and complex reasoning. Apple long sold its difference on two promises: privacy, and the vertical integration of silicon and software. The first still holds, on paper. The second was just buried in a single sentence.

Because all five models are, in Apple’s own words, built in collaboration with Google. Pre-training ran on Google’s TPU accelerators. And, most strikingly, AFM 3 Cloud Pro, the top of the range, runs on NVIDIA GPUs hosted in Google Cloud, with Apple having extended its Private Cloud Compute to that third-party infrastructure. Read that again: the most powerful model from the company that made autonomy its creed runs on a supplier’s silicon (NVIDIA) and in a direct competitor’s cloud (Google).

Let us be precise about what is actually new here, because the word dependency deserves to be weighed. Apple has always relied on third-party components, and no one bats an eye: the screens come partly from Samsung, the modems were long signed Qualcomm, the memory and storage come from a handful of Asian suppliers. The company built its industrial genius on assembly and orchestration as much as on pure design. The 2026 break, then, is not the outsourcing itself; it is where the outsourcing now sits. Until now, the third party supplied the panel or the radio chip; the brain, the operating system, the in-house silicon, the intelligence, all stayed stamped Apple. This time it is the very core of the flagship product, the AI, precisely where Apple had promised the most independence and on which it had staked its entire “designed by Apple in California” narrative, that has come to depend on a competitor. The slogan never said “assembled”; it said “designed.” Yet on what matters most for the decade ahead, the design is now done by two, and the cutting-edge execution is rented elsewhere.

We recognize here an oxymoron I have dissected before: that of a proclaimed sovereignty resting on someone else’s infrastructure. It was exactly the contradiction of Wero, the payment system sold as sovereign yet hosted on Amazon. It is also the dilemma of Mistral, which takes on debt for data centers while remaining captive to NVIDIA. Apple does not escape the rule; it confirms it on a spectacular scale. The flash sparsity on the device, brilliant as it is, ends up looking like a fig leaf: you are shown total command over the few billion active parameters in your pocket, the better to draw your eye away from the hundreds of billions running next door.

One technical detail is worth flagging for anyone tracking the trajectory of agentic AI: it is precisely the model entrusted to Google’s infrastructure that carries the autonomous tool-use workloads, the very ones I described as what moves AI from RAG to the agent. The frontier, meaning the part that counts most for the future, is the one Apple controls least. The innovation lives in the on-device constraint. The dependency lives in the horsepower.

Europe, a Market Watched From a Distance

That leaves the news that, seen from Saint-Marcellin, stings the most. On the same day, Apple unveiled Siri AI, a complete overhaul of the assistant built on this new generation of models. And in the same breath, the company announced that European Union users would not get it on iOS 27 and iPadOS 27 at launch, with no timeline for catching up. The stated reason: the Digital Markets Act, whose interoperability requirements would, by Apple’s account, force it to open its assistant to competitors at the cost of privacy and security.

Savor the irony. Apple is in effect telling Europeans: “We have built the most privacy-respecting on-device AI ever made, but your competition law keeps us from giving it to you.” And to drive the point home, the feature will remain available in the EU on macOS 27, visionOS 27, and watchOS 27, just not on iPhone or iPad, because iOS alone is designated a gatekeeper platform subject to the DMA’s interoperability obligations. Worse for our ecosystem: developers based in the EU will not even be able to test these Siri AI features for their own apps on iPhone and iPad.

We legislate for openness and reap exclusion. And along the way we deny our own toolmakers the means to build.

Does that absolve Apple? No. This sidelining is also a bargaining chip, a way of making users pay the price of a regulatory standoff, in the hope that they turn on Brussels rather than on Cupertino. Both camps are playing their hand on the back of the European consumer. But the net result is unambiguous: a continent where a growing share of what gets shown on stage never arrives, or arrives trimmed down, or arrives late. It is the same mechanism I described in writing about our shrinking technological sovereignty, somewhere between SCAF and IRIS²: we excel at regulating a playing field we no longer play on. And as long as Europe lacks champions of its own, which is to say as long as it does not embrace a sovereign AI built on local open source, it will remain a spectator dictating the rules of a match to which it is not invited.

The question is still worth asking, without doom-mongering: when so large a share of what is shown on stage never crosses our borders, what remains, for the European user, of the meaning of a global keynote? The risk is not that they stop inviting us; it is that, little by little, they stop thinking of us when they design the products. The gap does not widen out of hostility; it widens out of habit. And it is precisely against that habit that the responsibility falls to us to build an alternative, rather than merely lamenting the shop window.

The Takeaway

Apple has won the memory battle. AFM 3 Core Advanced is a gem of on-device engineering, an avowed heir to research begun in 2023, and it pushes back, for good, the limits of what a phone can run locally. Let us applaud it plainly, but without naivety: it is a physicist’s compromise as much as a stroke of genius.

Yet in the same motion, the company set down two retreats. The retreat from vertical integration, by entrusting its frontier to Google and NVIDIA. And, on our continent, an imposed retreat: that of a Europe shown the future the way one shows a shop window, behind glass, with nothing for sale. The Apple that once sculpted the chamfer of an iPhone 5s like a piece of fine jewelry did everything, controlled everything, showed everything. The Apple of 2026 still works wonders in your pocket, but the horsepower, now, is rented elsewhere. The show, more and more, is designed without us. What remains to be seen is whether we want to stay in the audience, or build a stage of our own.