Apple Machine Learning Research published reference benchmarks and principles for deploying Transformer models on the Apple Neural Engine via Core ML, replacing nn.Linear projections with 4D spatial convolutions (bc1s) and splitting execution across CPU, GPU, and Neural Engine compute units to cut latency and peak memory on mobile devices.
Apple Machine Learning Research
Also known as: Apple ML · Apple Foundation Models · MLX · OpenELM · FastVLM
Plain English
Apple's machine learning research and engineering division develops the frameworks, open research models, and foundation language and vision models that run across iOS, iPadOS, macOS, and Private Cloud Compute.
Technical Overview
Apple Machine Learning Research co-designs model architectures, low-bit Quantization, and execution runtimes (Core ML and the open-source MLX array framework) around Apple silicon's unified memory and Neural Engine. Its work spans transparent open-weight research artifacts (OpenELM, MM1, FastVLM) and the dual-tier Apple Foundation Models stack: compact on-device models paired with LoRA adapters and server-side Parallel-Track Mixture-of-Experts (PT-MoE) models served on Private Cloud Compute.
Timeline (6)
2022
Ecosystem2023
ReleaseApple open-sourced MLX, a NumPy- and PyTorch-inspired array framework engineered for Apple silicon's unified memory architecture so CPU and GPU operations share tensor buffers directly without host-to-device memory copies.
2024
ReleaseFollowing its MM1 multimodal pre-training study in March 2024, Apple released OpenELM (270M to 3B parameters) alongside complete CoreNet training logs, multiple Checkpoints, and MLX conversion utilities—using layer-wise scaling of attention heads and feed-forward dimensions to allocate parameters more efficiently across Transformer layers.
- Release
Apple introduced the Apple Intelligence Foundation Language Models, pairing an approximately 3B-parameter on-device model (using grouped-query Attention, low-bit palletization, and dynamically swapped LoRA adapters) with a larger server foundation model running on Apple silicon Private Cloud Compute.
2025
ReleaseApple detailed its 2025 foundation model upgrades and the Swift Foundation Models framework—combining 2-bit quantization-aware training and KV-cache sharing on-device with a Parallel-Track Mixture-of-Experts (PT-MoE) server backbone and the open-weight FastVLM (
FastViTHD) hybrid vision encoder for low-latency high-resolution visual understanding.2026
ReleaseApple introduced the third generation of Apple Foundation Models (AFM 3) alongside MLX optimizations for M5 GPU Neural Accelerators and Thunderbolt RDMA clusters, shipping AFM 3 Core and a sparse native multimodal AFM 3 Core Advanced on-device model (~20B total parameters activating 1B–4B per token across flash and active memory).
See also (11)
Open full graphApple Machine Learning Research