Apple

corporate

usa · Founded 1976 · IPO 1980 (NASDAQ: AAPL) · Market cap: $3.76T as of 2026-04

Apple's AI strategy centers on on-device + Private Cloud Compute rather than API-served models. Apple Intelligence, announced at WWDC 2024, integrates foundation models across iPhone, iPad, and Mac for writing tools, image generation, Siri enhancements, and cross-app actions.

The Apple Foundation Models (AFM) include a ~3B on-device model and larger server models, using a proprietary mixture-of-experts architecture. Apple published technical reports in 2024 and 2025 detailing the architecture, training, and safety measures. The company also contributes open research: MLX (25K stars), a NumPy-like ML framework optimized for Apple Silicon, and OpenELM, fully open efficient language models.

Apple's scaling science program spans three complementary groups: Afshin Dehghan leads vision tokenization research (AToken, FlexTok), Josh Susskind (Distinguished Engineer) leads multimodal architectures and scaling laws (NMM Scaling Laws, AIMv2), and Pierre Ablin leads training optimization (AdEMAMix, data mixture scaling laws). Samy Bengio (103K+ citations) leads ML research overall, with Amar Subramanya named VP of AI in December 2025.

Website Wikipedia HuggingFace GitHub
on-devicemultimodalopen-weight

People

News

Outputs (9)

AToken

paper

"A Unified Tokenizer for Vision." The first unified visual tokenizer achieving both high-fidelity reconstruction and semantic understanding across ima…

Scaling Laws for Optimal Data Mixtures

paper

Systematic method using scaling laws to determine the optimal data mixture for any target domain, replacing trial-and-error approaches. Predicts model…

Scaling Laws for Native Multimodal Models

paper

Extensive scaling laws study spanning 457 trained models examining native multimodal models (NMMs) trained from scratch on all modalities. Finds no in…

FlexTok

paper

"Resampling Images into 1D Token Sequences of Flexible Length." Projects 2D images into variable-length, ordered 1D token sequences (1 to 256 tokens f…

AIMv2

paper

"Multimodal Autoregressive Pre-training of Large Vision Encoders." Pairs a vision encoder with a multimodal decoder that generates both raw image patc…

AdEMAMix Optimizer

paper

"The AdEMAMix Optimizer: Better, Faster, Older." A modification of Adam using a mixture of two exponential moving averages to exploit older gradients…

Apple Foundation Models (AFM)

paper

Technical report detailing the foundation models powering Apple Intelligence. ~3B on-device model and larger server models running on Apple's Private…

OpenELM

model

Open Efficient Language Models. 270M to 3B parameter dense Transformers with layer-wise scaling (varying width per layer for parameter efficiency). Fu…

MLX

library

Array framework for machine learning on Apple Silicon. NumPy-like API with lazy computation, dynamic graph construction, and unified memory (arrays li…