Back to Gap Map Infrastructure / Classic ML & computer vision

timm (PyTorch Image Models)

open source / Overall score: 4.2(strong)

PyTorch library of image models, layers, optimizers, schedulers, augmentations and reference training and validation scripts, spanning more than 100 architecture families from ResNet and EfficientNet to ViT, ConvNeXt and MobileNetV4, with over 700 pretrained weights. Created by Ross Wightman and maintained under Hugging Face.

Openness

5 low confidence

5.0

license: Apache-2.0(OSI
source: public(github.com/huggingface/pytorch-image-models)
core features withheld: no — no paid tier or commercial edition of the library found

The library is Apache-2.0 and builds whole from the public repository, and its README and docs mention no paid tier or commercial license for it. The pretrained weights it loads are a different matter: each carries its source's terms, and some, such as Facebook's, are non-commercial.

https://raw.githubusercontent.com/huggingface/pytorch-image-models/HEAD/LICENSE recorded 2026-09-26
LICENSE: Apache License, Version 2.0, "Copyright 2019 Ross Wightman".
https://raw.githubusercontent.com/huggingface/pytorch-image-models/HEAD/README.md recorded 2026-09-26
README: "The code here is licensed Apache 2.0." and "The Facebook models have an explicit non-commercial license (CC-BY-NC 4.0, ...)"; no paid tier or commercial license for timm mentioned.
https://ungh.cc/repos/huggingface/pytorch-image-models/files/HEAD recorded 2026-09-26
Repository tree of 504 files (timm, tests, hfdocs, train.py, validate.py, inference.py) with no ee, enterprise, commercial or pro directory.

Adoption

5 high confidence

5.0

Measured on monthly PyPI downloads of timm only; Hugging Face downloads of the timm checkpoints are left out because that namespace also mirrors other vendors' weights.

https://pypistats.org/api/packages/timm/recent recorded 2026-09-26
last_month 10,328,111 downloads of timm

Capability

3 medium confidence

3.0

timm is the widest zoo of image backbones, with the training recipes to reproduce and fine-tune them. A new task still means training or fine-tuning on labeled images; it leaves the choice of architecture to the user.

https://huggingface.co/docs/timm/index recorded 2026-09-26
Docs: "timm is a library containing SOTA computer vision models, layers, utilities, optimizers, schedulers, data-loaders, augmentations, and training/evaluation scripts. It comes packaged with >700 pretrained models".
https://raw.githubusercontent.com/huggingface/pytorch-image-models/HEAD/README.md recorded 2026-09-26
README Models list of 106 architecture families (BEiT, ConvNeXt, CoAtNet and others), with sections on optimizers, augmentations and regularization.

Verified 2026-09-26