Apple Foundation Models: The 3B AI Powering Apple Intelligence

3B parameters, native on iOS 26 / iPadOS 26 / macOS 26, free Foundation Models framework — but Apple-only and closed-weight.

Last reviewed: May 2026
Parameters3 B
Size (quantized)Not disclosed
Context lengthNot disclosed
Modalitytext+vision
Licenseapple-proprietary
Min RAM8 GB
VersionAFM-on-device (iOS 26)
Released2025-09

What is it?

Apple Foundation Models (AFM) is the family of large language models that powers Apple Intelligence. Apple ships two variants: a 3 billion parameter on-device model that runs entirely on the user’s device, and a larger server-based model used for more complex queries via Private Cloud Compute. With iOS 26 / iPadOS 26 / macOS 26, Apple opened the on-device model to third-party developers through the Foundation Models framework — directly callable from Swift, free, and with no per-request cost.

Core specs at a glance

(See spec card above — populated from structured data. Apple has not publicly disclosed the quantized size or context window length, so those fields display as “Not disclosed”.)

What devices can run it?

AFM-on-device requires Apple Intelligence support: iPhone 15 Pro / 15 Pro Max / 16 family on the phone side, iPad with M1 chip or newer, and Mac running Apple silicon. The runtime is the Apple Neural Engine plus Metal acceleration, transparently managed by the OS. Older devices simply do not get Apple Intelligence features. There is no fallback to a smaller model — it is all-or-nothing per device.

Strengths and limitations

Strengths. Zero install — every supported device already has the model. Free for developers via the FoundationModels framework. Tight Swift integration with guided generation, structured output, and tool calling. Excellent privacy story: inference is local, optionally Private Cloud Compute for harder queries with cryptographic verification.

Limitations. Apple-only. Closed weights — you cannot inspect, fine-tune, or self-host. Many spec details (quantization, context length) remain undisclosed by Apple. Cross-platform apps still need a second model for non-Apple devices, doubling QA and content surface area.

When to choose it (and when not to)

Choose AFM-on-device if: your app is exclusive to Apple platforms; you want zero-install AI; you value privacy guarantees backed by Apple’s full-stack control; you want the simplest developer experience for tasks like content tagging, summarization, or structured generation.

Skip it if: you need to ship beyond Apple (Android, Windows, Linux); you need open weights for fine-tuning, distillation, or compliance audits; you need known specs (Gemma 4 publishes everything Apple keeps private).

How it compares to similar on-device models

Closest alternatives are Gemma 4 E2B (2.3B effective, open Apache 2.0, runs anywhere) and Microsoft Phi-4-multimodal (5.6B, MIT-licensed, also adds audio modality). Apple’s strength is the integration story; the open competitors win on portability and inspectability. For a side-by-side, see the leaderboard.

In a real Cove app

Cove Travel and Cove Voice use Gemma 4 across iOS and Android because we need the same model to run on both platforms. AFM-on-device is the model we’d use if Cove were Apple-only — the integration story is unbeatable when your audience is 100% iPhone / iPad. The thinking model behind Cove’s privacy promises (“inference never leaves your device, by architecture”) aligns directly with what Apple Foundation Models stands for.

See it in a real Cove app

FAQ

What is Apple Foundation Models?

Apple Foundation Models (AFM) is the family of LLMs powering Apple Intelligence. Developers can call the on-device 3B-parameter variant directly from Swift through the FoundationModels framework, free of charge, with full privacy guarantees because inference never leaves the device.

What devices does AFM-on-device run on?

Any Apple Intelligence-compatible device: iPhone 15 Pro / 15 Pro Max / 16 family, iPad with M1 chip or newer, and Mac with Apple silicon (M1, M2, M3, M4). Older Apple devices fall back to cloud or simply lack Apple Intelligence features.

Is Apple Foundation Models open source?

No. The weights are proprietary and not downloadable. You can only use the model through Apple's FoundationModels framework on Apple devices. This is the major trade-off versus Gemma 4, Phi, or Qwen, which all ship open weights you can run anywhere.

How big is the on-device model?

Apple has not published the quantized size or context window. Apple's 2024 reports described the on-device model as pruned from a 6.4B base; the 2025 tech report describes a different distillation approach using a sparse-upcycled 64-expert MoE teacher derived from a ~3B model. Either way, the deployed on-device model has ~3B effective parameters.

How does it compare to Gemma 4 or Phi-4?

AFM-on-device is competitive with Qwen-2.5-3B and Gemma-3n per Apple's own benchmarks, and gives a free, integrated developer experience on Apple platforms. Gemma 4 and Phi-4-multimodal beat it on cross-platform reach (Android + Windows + Linux) and openness, but require more setup. Pick AFM if your users are 100% Apple.

Citations