The Open-Source AI Stack
All models

Models · apple-foundation

Apple On-Device Foundation Model (2025)

Proprietary Apple · 2025-06-09 · Proprietary

Apple Intelligence's on-device foundation model, announced WWDC 2025 on June 9 and shipped in iOS 26. About 3B parameters with KV-cache sharing across blocks (37.5 percent KV cache reduction) and 2-bit quantization-aware training, paired with a server-side Parallel-Track Mixture-of-Experts model on Private Cloud Compute. Foundation Models framework opened direct model access to developers.

Architecture

tokens in Embedding vocab not disclosed × N layers Attention (not disclosed) Position encoding not disclosed context not disclosed Dense MLP SwiGLU activation (standard) 3B active params Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
dense
Total params
3B
Active params
3B
Context window
not verified
Attention
unknown
Position encoding
unknown
Post-training
sft, rlhf
OSI-approved
no
Data released
no
Training code
not released
QAT
2-bit (official)

Available quantizations

QAT 2-bit Trained to be quantized: quantization-aware training, with official checkpoints (2025-07-17).

Decoder weights compressed to 2 bits per weight via QAT (embedding table 4-bit, KV cache 8-bit); LoRA adapters recover quality.

source ↗

None. The weights are not distributed, so there are no public quantizations.

Notable innovations

  • · 2-bit quantization-aware training for on-device
  • · KV-cache sharing between blocks (37.5 percent reduction)
  • · Foundation Models framework for app developers
  • · Parallel-Track MoE server companion

Lineage

Apple's first developer-accessible foundation model; ships with iOS 26.

Sources