Apple drops 20 on-device Core ML models on Hugging Face

Apple just released 20 new Core ML models and four datasets on Hugging Face, giving developers ready-to-use models optimized to run locally on Apple hardware without setting up cloud infrastructure.

What the models can do

The collection covers core computer vision tasks. Key additions include FastViT for fast image classification, DepthAnything for monocular depth estimation, and DETR for semantic segmentation.

Hugging Face CEO Clement Delangue noted that these models power practical, everyday app features. Think of an app that instantly cuts out photo backgrounds or identifies real-world objects through the camera and names them in another language.

Alongside the 20 models, Apple uploaded four datasets to help developers fine-tune and evaluate their own implementations.

Tuned strictly for on-device performance

Every model in this drop is compiled in Core ML format, tailored specifically to take advantage of Apple Silicon while keeping memory and power usage low.

They run entirely on the user's device with zero network connection required. That brings two major advantages: near-instant local inference without cloud latency, and absolute data privacy since no photos or personal files ever leave the hardware.

Developers can find and download all 20 models and datasets directly from Apple's repository on Hugging Face.

Why it matters

For iOS and macOS developers, this eliminates the tedious process of converting PyTorch checkpoints to Core ML and tuning them for mobile chips. You get drop-in assets ready for Xcode.

Running basic vision tasks locally cuts ongoing cloud API bills down to zero while meeting strict privacy expectations. If you build for the Apple ecosystem and need image segmentation or depth detection, we think this drop is well worth testing right away.