Apple’s third-generation Basis Fashions defined

In the course of the WWDC26 keynote, Apple introduced its third era of Apple Basis Fashions (AFM), comprising 5 fashions, a few of that are native, a few of that are cloud-based, and certainly one of which lives in Google’s servers operating on Nvidia chips. Right here’s a breakdown of how that may work.

A little bit of background

When Apple first introduced its basis fashions in 2024, the lineup included an on-device language mannequin with roughly 3 billion parameters, and “a bigger server-based language mannequin accessible with Personal Cloud Compute and operating on Apple silicon servers,” as the corporate put it on the time.

Personal Cloud Compute was an formidable enterprise, because it aimed to ship cloud-based AI capabilities whereas preserving the identical privateness ensures customers anticipate from on-device processing.

For that reason, conserving all the things in-house was important. Personal Cloud Compute ran in Apple information facilities, on servers powered by Apple silicon. Even so, its privateness ensures may very well be independently verified by third-party safety researchers.

Nonetheless, as Apple struggled to get its AI aspirations off the bottom, the corporate partnered with Google to make use of Gemini because the spine of its new AI efforts, the outcomes of which it introduced earlier this week through the WWDC26 keynote.

Apple’s new basis fashions

The third era of AFMs contains 5 fashions: AFM 3 Core and AFM 3 Code Superior, that are on-device fashions, and AFM Cloud, ADM 3 Cloud (Picture), and AFM 3 Cloud Professional, that are server-based. The D in ADM 3 Cloud (Picture) stands for diffusion, a know-how we’ve lined previously right here.

Aside from AFM 3 Cloud Professional, all different fashions had been constructed to run on Apple silicon units. AFM 3 Cloud Professional, in the meantime, runs on NVIDIA GPUs hosted in Google Cloud.

This was made potential afer Apple prolonged its Personal Cloud Compute structure to third-party infrastructure for the primary time, “whereas sustaining Apple’s highly effective safety and privateness protections,” in line with the corporate.

As for the fashions themselves, right here’s a breakdown of every one, as defined by Apple:

• AFM 3 Core, the subsequent era of our 3-billion-parameter dense mannequin that delivers a step up in high quality.
• AFM 3 Core Superior, our strongest on-device mannequin. It’s natively multimodal, enabling useful options like expressive voices and higher-accuracy dictation. Constructed on cutting-edge Apple analysis, this 20-billion-parameter mannequin makes use of a sparse structure, activating simply 1 to 4 billion parameters at a time relying on the request. AFM 3 Core Superior is unlocked by and optimized for our most succesful Apple silicon techniques.
• AFM 3 Cloud, our server-side workhorse, optimized for velocity, effectivity, and efficiency.
• ADM 3 Cloud (Picture), for picture era and enhancing, which unlocks superior photo-editing instruments, the all-new Picture Playground, and extra.
• AFM 3 Cloud Professional, our most succesful server-based mannequin, which powers our most demanding use instances, like agentic device use and sophisticated reasoning.

The highlights listed below are AFM 3 Core Superior and AFM 3 Cloud Professional.

Starting with AFM 3 Core Superior, it packs 20 billion parameters into an on-device mannequin, which is not any small feat. Most on-device fashions aimed toward most of the people have a tendency to remain within the low-single-digit billions of parameters.

To make AFM 3 Core Superior run nicely, Apple used a sparse structure that prompts as much as 4 billion parameters at a time, relying on the immediate, reasonably than a dense structure that would want to maintain all 20 billion parameters energetic for each request.

Though conceptually much like the Combination of Consultants method, this selective activation depends on a way Apple invented and detailed within the fascinating research Instruction-Following Pruning for Massive Language Fashions launched a yr in the past.

As for AFM 3 Cloud Professional, that is the one which runs on an exterior infrastructure. You’ll be able to learn among the technical particulars of this enlargement in this text printed on Apple’s Safety weblog earlier this week, however right here’s crucial half:

On this basis, Apple and Google collaborated to construct capabilities that go far past a conventional confidential computing deployment:

• We don’t rely solely on confidential computing applied sciences to mitigate assaults that leverage privileged entry exterior of a confidential VM, together with side-channel assaults. We contemplate each element — from firmware by means of the host and visitor OS stacks to utility code — to be a part of our trusted computing base, topic to our verifiable transparency and no-privileged-access ensures.
• To mitigate the chance of provide chain assaults, we preserve a cryptographically verifiable, append-only ledger of all Google Cloud {hardware} that’s a part of the PCC fleet. For parts that may very well be abused to exfiltrate consumer information if compromised, our software program attestation is rooted in at the least two separate roots of belief from unbiased distributors.
• Even when deployed with confidential computing, we imagine the inference stack have to be designed with privateness and safety from the beginning. PCC on Google Cloud leverages lots of the similar architectural safety patterns as PCC on Apple silicon to implement these layered protections: preliminary community information parsing for every request occurs in a devoted course of inside its personal namespace, shared inference software program is recycled with a brief time-to-live period, and attested keys are held in a separate, devoted confidential VM remoted from exterior inputs.

In its Machine Studying Analysis weblog, Apple says that every one 5 fashions “shared a standard preliminary basis earlier than specializing for his or her respective architectures and use instances, including multimodal capabilities like audio, picture understanding, long-context reasoning, and high-quality visible era.”

The corporate provides that, to coach these fashions, it used “a combination of knowledge that features publicly accessible data, information licensed or bought from third events, open-sourced information, information obtained by means of devoted research, and artificial information.” Apple additionally stresses that the coaching course of didn’t embrace consumer information or interactions and that net publishers can choose out of basis mannequin coaching.

The outcomes

Apple says it carried out in depth human evaluations of its third-generation basis fashions, with in-house reviewers grading responses throughout classes reminiscent of instruction following, truthfulness, presentation, and picture understanding.

Fashions had been evaluated towards their predecessors (when relevant), and you may see among the outcomes under:

Fraction of most popular responses in side-by-side human evaluations of basic textual content capabilities, evaluating AFM 3 Core and AFM 3 Cloud towards our earlier era of fashions. Outcomes are introduced throughout 4 distinct locale teams to reveal constant efficiency throughout worldwide variants. “English” represents our world English analysis set, whereas “PFIGSCJK”, “DNNSTV” and “AFIHHMPRTU” characterize our remaining supported world locales.

Fraction of most popular responses in side-by-side human evaluations of picture understanding capabilities in English. The outcomes examine AFM 3 Core and AFM 3 Cloud towards their 2025 predecessors.

Fraction of most popular responses in side-by-side human evaluations for dictation duties. The outcomes examine AFM 3 Core Superior towards Apple’s current manufacturing dictation system throughout seven high quality dimensions. AFM 3 Core Superior demonstrates a optimistic win charge in general high quality, with desire extending persistently throughout all particular person formatting and comprehension dimensions.

For an excellent deeper dive into the third-gen Apple Basis Fashions, comply with this hyperlink.

Value testing on Amazon