Apple Has 5 New AI Models, Distilled from Google’s Gemini

Apple unveiled the third generation of its Apple Foundation Models (AFM) at WWDC 2026 — a family of five models that, for the first time, is built openly on Google’s Gemini technology. From the 3-billion-parameter model on the iPhone to the reasoning powerhouse in the cloud, Apple is promising a generational leap.

It’s the deal that turned Apple’s AI strategy on its head: on January 12, 2026, Apple and Google announced a multi-year partnership that moves Google’s Gemini models to the center of the next foundation-model generation. “After careful evaluation, we determined that Google’s technology provides the most capable foundation for the Apple Foundation Models,” the joint statement said. Bloomberg reporter Mark Gurman put the scope at roughly one billion US dollars per year — for access to a custom Gemini model with a reported 1.2 trillion parameters.

At WWDC 2026, it became clear what had been built on this foundation. Apple still speaks of “Apple Foundation Models” developed “in collaboration with Google” — so the branding stays Apple, while the underpinnings come from Gemini. Training was done on Google’s cloud TPUs, and hosting runs through Private Cloud Compute. Crucial to the privacy story Apple is determined to keep: user data would “never be stored or shared — not even with Apple,” and Google confirmed that it receives no Apple user data.

Two models for the device

The core of on-device intelligence is formed by two models that run directly on iPhone, iPad, and Mac.

AFM 3 Core is the next generation of the familiar dense 3-billion-parameter model and the all-rounder for everyday tasks. In Apple’s own side-by-side tests with human evaluators, it was preferred over its predecessor in 45.6 percent of cases on general text tasks — the 2025 baseline managed only 23.3 percent. The new generation also came out ahead on image understanding.

AFM 3 Core Advanced is the real highlight on the device — and technically the most exciting model in the family. It is natively multimodal and powers features such as expressive voices and a more precise dictation function. At 20 billion parameters, it would actually be too large for most smartphones, but Apple uses a new architecture: depending on the request, only 1 to 4 billion parameters are active at any one time.

The trick behind this is called Instruction-Following Pruning (IFP), a technique developed by Apple researchers. Instead of forcing the entire model into fast working memory (DRAM), it resides completely in flash storage (NAND). Because the bandwidth between NAND and DRAM is too slow to swap weights token by token, the model makes its routing decision per prompt: a lightweight block selects a fixed set of “experts” at the outset and combines them with permanently active “shared experts.” This makes it possible to scale the model size far beyond the usual DRAM limits — with minimal latency. AFM 3 Core Advanced is unlocked only on Apple’s most powerful chips.

How big the leap is in listening and speaking is shown by the feature tests: in speech synthesis, AFM 3 Core Advanced reached a Mean Opinion Score of 4.15 versus 3.87 for the previous production system — and on colloquial text even 4.24 to 3.82. In dictation, overall quality was preferred in 44.7 percent of cases, the old system in only 17.6 percent.

Three models in the cloud

Three further models for more demanding tasks run in Private Cloud Compute.

AFM 3 Cloud is the server-side workhorse, optimized for speed, efficiency, and multimodal reasoning. Here Apple built on the Parallel-Track Mixture-of-Experts architecture (PT-MoE) introduced the previous year and refined it to stabilize training and improve recall of information within the context window. The generational leap is clear: in direct comparisons, AFM 3 Cloud was preferred on text tasks in 64.7 percent of cases, the 2025 server model in only 8.7 percent. On image understanding, the figures stand at 37.8 to 9.6 percent.

ADM 3 Cloud (Image) is the model for image generation and editing. It powers the advanced photo tools, the overhauled Image Playground, and Genmoji. Apple emphasizes strong controllability and parameter efficiency: the model generalizes across different aspect ratios and resolutions and uses specialized adapters for individual editing features — such as “Spatial Reframing” in the Photos app.

AFM 3 Cloud Pro is the most powerful server model for the most demanding use cases: agentic tool use and complex reasoning. Compared with AFM 3 Cloud, it adds another roughly 10 percent in response quality for text and 14 percent in image understanding; on math tasks it gains 14 percent. Notably, as the only one of the five models, AFM 3 Cloud Pro was optimized not for Apple Silicon but for NVIDIA GPUs. To this end, Apple — together with Google and NVIDIA — extended Private Cloud Compute to NVIDIA GPUs in Google Cloud, with what Apple says are identical privacy guarantees.

What this means for users

In concrete terms, the five models are meant to enable a completely new Siri, advanced photo editing, an upgraded Image Playground, and significantly more expressive voices. Precisely those Siri functions — personal context, screen awareness, and the execution of multi-step app actions — that Apple had already announced back in 2024 and then repeatedly postponed are now supposed to arrive at last, on the Gemini foundation.

Apple explicitly classifies the figures presented as a snapshot during the beta phase; it announced a technical report with updated benchmarks for later in the summer. Independent benchmarks and tests do not yet exist — no corresponding evaluations of the AFMs can be found on either Arena.ai or Artificial Analysis.

For Apple, the move is a remarkable balancing act: the company that relied for years on its own models and had maneuvered itself behind in generative AI is now licensing the technology from, of all rivals, Android-maker Google — yet consistently sells the underpinnings as its own “Apple Foundation Models” and puts privacy, via Private Cloud Compute, front and center. Whether this narrative holds up in everyday use will become clear once the first Gemini-driven Siri features actually land on devices.