timm

Nearly every modern computer vision baseline lives in timm, Ross Wightman's collection of PyTorch image models that covers more than 100 architecture families and over 1,000 pretrained weight configurations, from ResNet, EfficientNet, and ConvNeXt to ViT, Swin, MobileNetV4, and recent additions such as DINOv3 and SigLIP-2 backbones. Beyond model definitions, the library ships reference training, validation, and inference scripts, a long list of optimizers running from AdamW and LAMB to Muon, Kron, and MARS, augmentation strategies including Mixup, CutMix, and RandAugment, learning rate schedulers, and consistent feature extraction APIs that make any backbone usable in detection or segmentation pipelines. A single pip install timm and a call to timm.create_model with pretrained weights pulled from the Hugging Face Hub is enough to start, on CPU or GPU. The project is Apache 2.0 licensed, hosted under the Hugging Face organization, and past 37,000 GitHub stars, with version 1.0.28 shipping in July 2026. It is the de facto backbone library for computer vision research and production transfer learning alike.