# Apple Foundation Models: the AFM 3 family explained

Apple’s third-generation foundation models (June 2026): AFM 3 Core (3B on-device), AFM 3 Core Advanced (20B sparse, multimodal), AFM 3 Cloud, ADM 3 Cloud and AFM 3 Cloud Pro. What each does and how the Foundation Models framework exposes them.

Canonical URL: https://aiidelist.com/blog/apple-foundation-models

Language: en

Published: 2026-09-24

Updated: 2026-09-24

APPLE FOUNDATION MODELS · AFM 3

On 8 June 2026 Apple introduced its third generation of foundation models. Two run on-device—AFM 3 Core and the sparse, multimodal AFM 3 Core Advanced—and three run on Private Cloud Compute. Developers reach them through the Foundation Models framework rather than by downloading weights.

[Foundation Models framework guide](/blog/apple-foundation-models-framework)[Apple Research announcement](https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models)[Foundation Models vs Core ML vs MLX](/blog/foundation-models-vs-core-ml-vs-mlx)

## How AFM 3 differs from FastVLM and open models

FastVLM, MobileCLIP2 and Depth Pro are open research models you download and run anywhere—browser, Python, MLX. Apple Foundation Models are not distributed as weights: they ship with the operating system and are used through the Foundation Models framework on Apple Intelligence devices. That means zero download for your users and system-level privacy, but also that Apple picks the model and the platform.

## Core Advanced: a 20B model that fits in a phone

Apple describes AFM 3 Core Advanced as a sparse architecture whose full weights live in NAND flash. Shared experts stay resident while routed experts are loaded on demand, and an instruction-following pruning scheme lets the model activate only 1–4B parameters per request. The result is a multimodal model with far more capacity than the memory it occupies.

## Using the models as a developer

The Foundation Models framework gives you a LanguageModelSession with guided generation into Swift types, tool calling, streaming and—since WWDC26—image input, Dynamic Profiles and a LanguageModel protocol that also accepts Private Cloud Compute, MLX and third-party providers. Python and a CLI are available for scripting and evaluation.

## Questions

**Can I download AFM 3 weights?**

No. Apple Foundation Models are accessed through the framework on supported devices and through Private Cloud Compute. For downloadable Apple models, see FastVLM, MobileCLIP2, Depth Pro, AIMv2 and SHARP.

**Which devices run AFM 3 Core Advanced?**

Apple ties model availability to Apple Intelligence support and device generation; check SystemLanguageModel availability at runtime instead of assuming a model tier.

## At a glance

**AFM 3 Core**

Dense ≈3B-parameter on-device model; the default for the Foundation Models framework

**AFM 3 Core Advanced**

≈20B sparse model activating 1–4B parameters per request; natively multimodal; weights stored in flash and loaded on demand

**AFM 3 Cloud · ADM 3 Cloud · AFM 3 Cloud Pro**

Private Cloud Compute models for speed, image generation/editing (Genmoji) and complex reasoning with agentic tool use

**Modalities**

All models are multimodal; Apple reports gains in text quality, image understanding and speech synthesis

**Training data**

Public, licensed and synthetic data; Apple states no private user data is used

Checked 2026-09-17

## Official sources

- [Introducing the Third Generation of Apple’s Foundation Models](https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models)
- [WWDC26 Machine Learning guide](https://developer.apple.com/wwdc26/guides/machine-learning/)

## Related

- [FastVLM on iPhone](/blog/fastvlm-ios-coreml)
- [All Apple models](/blog/apple-ai-models)
- [All Apple models](/blog/apple-ai-models)