Apple: the third generation of Foundation Models brings five models and dramatic quality improvements

Five models for device and cloud

Apple published the third generation of Apple Foundation Models (AFM 3) on June 8, 2026 — an integrated family of five models designed for on-device and cloud operation. The new generation replaces the previous second-generation AFM models and introduces an expanded architectural range. Each model covers a distinct technical and market domain.

AFM 3 Core (3 billion parameters) runs directly on iPhone, iPad, and Mac devices with Apple Silicon chips, without sending data to external servers. AFM 3 Core Advanced carries 20 billion parameters and uses a sparse Mixture of Experts (MoE) architecture. AFM 3 Cloud is a server-side model optimized for speed and operational efficiency. ADM 3 Cloud is specialized exclusively for image generation and editing. AFM 3 Cloud Pro runs on NVIDIA GPUs and is intended for complex reasoning that exceeds the capacity of on-device models.

What does the sparse MoE architecture mean for on-device AI?

AFM 3 Core Advanced stores all 20 billion parameters in the device’s flash memory. The technique called ‘Instruction-Following Pruning’ selectively loads only 1 to 4 billion active (‘expert’) parameters into DRAM for each incoming request. The remaining weights stay inactive in flash storage until needed for processing. The result is elasticity in inference: the model adjusts its memory load to the complexity of the task without constant pressure on operating memory. Previous generations of Apple Foundation Models did not use a MoE architecture — AFM 3 Core Advanced marks the first implementation of that technique in Apple’s on-device ecosystem.

Users prefer AFM 3 across all categories

An internal benchmark based on user preferences shows consistent gains for the new generation across all tested segments:

• AFM 3 Core: 45.6% of users prefer the new model, versus 23.3% who preferred the previous version.
• AFM 3 Cloud: 64.7% prefer the new model, versus 8.7% for the previous generation — a relative gain of 36% in user satisfaction.

The text-to-speech (TTS) component reached a Mean Opinion Score of 4.15, an improvement over the baseline value of 3.87. Dictation — automatic speech-to-text conversion — records 44.7% preference among respondents for overall output quality, indicating measurable progress in natural speech understanding.

Privacy and collaboration with Google

Apple explicitly confirmed that AFM 3 was not trained on user data or the content of user communications. The company respects opt-out requests from content publishers whose material may not be used for model training. Part of the server infrastructure was developed in technical collaboration with Google for optimization on specific hardware platforms — Apple states that this collaboration does not affect the model training process or the user privacy policy.

The third generation of AFM continues Apple’s ‘Private AI’ approach: on-device processing for sensitive tasks, selective escalation to the cloud for more complex requests, with permanent protection of user data from inclusion in training sets.