PyTorch Image Models — universally known as timm — is the largest open-source collection of image encoders and backbones for PyTorch. Created by Ross Wightman and now maintained with Hugging Face, it has earned more than 37,000 stars on GitHub and is updated constantly, with its latest changes in 2026. For computer-vision work in PyTorch, timm is the default toolbox.
The library provides implementations and pretrained weights for hundreds of architectures — ResNet, EfficientNet, Vision Transformers, ConvNeXt, Swin Transformer, MobileNet, MaxViT and many more — behind one consistent API. Training, evaluation, inference and export scripts share the same interface, so switching from one backbone to another is a one-line change. timm also supplies the surrounding machinery serious training needs: data augmentation such as RandAugment and Mixup, modern optimizers and schedulers, and distributed-training support, which is why it underpins countless research papers, Kaggle solutions and production vision systems. Most models can also be loaded straight from the Hugging Face Hub by name, and timm backbones power a large share of the vision models published there. For practitioners, that ubiquity means examples, pretrained checkpoints and community knowledge exist for nearly any architecture you might choose.
timm is completely free and open source under the Apache-2.0 license. Install it from PyPI and use it in research or commercial products; there is no paid tier of any kind.
timm is for computer-vision researchers, ML engineers and Kaggle competitors who need proven image models with reliable training recipes — without reimplementing every architecture from its paper.