AI Potluck
Contribute
Back to Gap Map Infrastructure / Model hubs & registries

PyTorch Hub

PyTorch Foundation (Linux Foundation)
source available / Overall score: n/a

Index of pretrained PyTorch models at pytorch.org/hub. A publisher adds a hubconf.py to its own GitHub repository and opens a pull request to pytorch/hub with a model page; once merged, the model is listed on the site and torch.hub.load fetches the code from the publisher's repository and the weights from wherever the publisher hosts them. The hub stores no weights itself. It is listed on the PyTorch Foundation's pytorch.org, and the loader ships inside PyTorch.

The submission repository has had no push since April 2024 and the site still calls the hub a beta release. The repository carries no license grant, so its pages can be read but not freely reused; the torch.hub loader is part of pytorch.

Openness

2 high confidence
2.0
license
none-declared(no LICENSE file in the 146-path tree and no license field in package.json
source
public(github.com/pytorch/hub, model pages and CI scripts)

The model pages, tag list and test scripts are all public, but the repository grants no license, and its contributing guide points at a LICENSE file that does not exist. All rights stay with the authors, so the hub can be read but not freely reused. No license is declared anywhere the publisher could declare one, so the license tier is unstated.

Adoption

not assessed

The index publishes no load or download counts, and torch.hub ships inside every PyTorch install, so PyTorch's own downloads measure the framework rather than the hub. No reading was recorded.

  • https://pytorch.org/hub/ recorded 2026-09-27

    Model index grouped by type (Audio 8, Generative 2, Nlp 3, Scriptable 21, Vision 41) with no download, pull or usage counters on the page

Capability

1 high confidence
1.0

Like the MONAI zoo, entries arrive by a reviewed pull request, but the hub keeps only a page per model: the code stays in the publisher's repository and the README says the hub does not host pretrained weights. That makes it an index over other people's hosting rather than a collection it stores and serves, which is the least any hub in this category does.

  • https://pytorch.org/hub/ recorded 2026-09-27

    "Discover and publish models to a pre-trained model repository designed for research exploration." "This is a beta release"; 75 type tags across the listed models

  • https://ungh.cc/repos/pytorch/hub/files/master/README.md recorded 2026-09-27

    "We accept submission to PyTorch hub through PR in `hub` repo." "Add a `hubconf.py` in your repo". "Currently we don't support hosting pretrained weights, users with pretrained weights need to host them properly themselves."

Verified 2026-09-27