OpenFold

weights

open(Apache-2.0 OpenFold3 checkpoints behind an auto-approved form, and OpenFold's own AlphaFold 2-reproduction parameters fetched by repository scripts)

data

open(OpenProteinSet, some 400,000 MSAs and PDB70 template hit files, on the AWS Registry of Open Data)

code

open(train_openfold.py plus a documented end-to-end training guide, not inference alone)

license

Apache-2.0(OSI)

OpenFold was built to be retrained from the start: the AlphaFold-2 reproduction ships a full training corpus (OpenProteinSet, on AWS's Registry of Open Data), a training pipeline (train_openfold.py and its guide), and the resulting weights, all under Apache-2.0. OpenFold can also load DeepMind's original AlphaFold 2 parameters, released separately under CC-BY-4.0. OpenFold3-preview ships checkpoints without a matching corpus or pipeline, so its openness reading depends on that AlphaFold-2 evidence.

• https://huggingface.co/api/models/OpenFold/OpenFold3 recorded 2026-10-06

  license: apache-2.0; private: false; gated: auto with the prompt "The OpenFold3-preview model is released under Apache 2.0 license. You will automatically get access to the model after answering the following simple questions"; four checkpoints/*.pt files present.

• https://openfold.readthedocs.io/en/latest/Training_OpenFold.html recorded 2026-10-06

  A training guide covering how to train an OpenFold model from the RODA dataset and how to fine-tune it, alongside a separate page on setting up the PDB training set.

• https://raw.githubusercontent.com/aqlaboratory/openfold/main/docs/source/original_readme.md recorded 2026-10-06

  "We've publicly released model weights and our training data - some 400,000 MSAs and PDB70 template hit files - under a permissive license. Model weights are available via scripts in this repository while the MSAs are hosted by the Registry of Open Data on AWS."

• https://raw.githubusercontent.com/aqlaboratory/openfold/main/LICENSE recorded 2026-10-06

  Apache License Version 2.0 body, matching the license the GitHub API reports for the repository.

• https://registry.opendata.aws/openfold/ recorded 2026-10-06

  OpenProteinSet listed on the Registry of Open Data on AWS as a public dataset, cited as OpenFold's training data.

The gated OpenFold3 endpoint reports zero downloads, which measures the gate rather than actual reach, so GitHub stars are used as the fallback signal instead.

• https://api.github.com/repos/aqlaboratory/openfold recorded 2026-09-10

  stargazers_count: 3423; archived: false; fork: false.

• https://huggingface.co/api/models/OpenFold/OpenFold3 recorded 2026-09-10

  downloads: 0 against likes: 44, on a gated: auto repository.

Two attributes, and the adaptation one is the whole point of the project - the README's first line is that this is a trainable reproduction, and the docs carry a training guide rather than a fine-tuning notebook bolted onto an inference release.

• https://raw.githubusercontent.com/aqlaboratory/openfold/main/docs/source/original_readme.md recorded 2026-09-10

  Training data is MSAs and PDB70 template hit files alongside sequence; the repository ships train_openfold.py beside run_pretrained_openfold.py.

• https://raw.githubusercontent.com/aqlaboratory/openfold/main/README.md recorded 2026-09-10

  "A faithful but trainable PyTorch reproduction of DeepMind's AlphaFold 2", with docs carrying instructions for installation and model inference and training.