Boltz
Jeremy WohlwendBoltz is a family of biomolecular interaction models. Boltz-1 was the first fully open-source model to approach AlphaFold 3 accuracy, and Boltz-2 jointly models complex structure and binding affinity, reporting a log10(IC50) value alongside a binder-versus-decoy probability. It runs from a YAML specification of the complex, optionally pulling MSAs from a ColabFold server.
Openness
5 high confidence- weights
- open(MIT weights downloaded by the boltz package, covering Boltz-1 and Boltz-2)
- data
- open(pre-processed RCSB and OpenFold structures and MSAs published for download, with instructions for reprocessing raw data)
- code
- open(training script and structure and confidence model configs under scripts/train, alongside a training guide)
- license
- MIT(OSI)
Boltz-1's full release - training corpus, preprocessing pipeline, training configs and MIT-licensed weights - is what makes this an open-source rather than open-weights entry. Boltz-2's equivalents are announced but not yet published, so the runnable training pipeline trails the model most users would want to retrain; the MIT license and weights, though, cover both versions.
- https://raw.githubusercontent.com/jwohlwend/boltz/main/docs/training.md recorded 2026-10-13
Download links for the pre-processed RCSB and OpenFold structures and MSAs and the ligand symmetry file, instructions for reprocessing raw data, and template training configs for the structure and confidence models under scripts/train/configs; Boltz-2 training information is marked coming soon.
- https://raw.githubusercontent.com/jwohlwend/boltz/main/LICENSE recorded 2026-10-13
MIT License body, Copyright (c) 2024 Jeremy Wohlwend, Gabriele Corso, Saro Passaro.
- https://raw.githubusercontent.com/jwohlwend/boltz/main/README.md recorded 2026-10-13
"All the code and weights are provided under MIT license, making them freely available for both academic and commercial uses"; a Training section pointing at docs/training.md for retraining.
Adoption
2 high confidenceAdoption is measured from PyPI downloads of boltz, the declared package that installs the model and fetches its checkpoints, because PyPI takes precedence over Hugging Face and GitHub stars whenever a package is declared. The Hub repositories boltz-community/boltz-1 and boltz-community/boltz-2 report zero downloads because the package fetches checkpoint files directly rather than through the Hub API that download counters measure.
- https://api.github.com/repos/jwohlwend/boltz recorded 2026-09-10
stargazers_count: 4202; archived: false; fork: false — recorded so the superseded stars signal stays visible.
- https://huggingface.co/api/models/boltz-community/boltz-2 recorded 2026-09-10
downloads: 0; likes: 16; gated: false; license mit. The Hub artifact exists and reports zero downloads.
Capability
3 high confidenceTwo attributes. Adding binding affinity to structure prediction widens what Boltz-2 outputs without adding a scale or a design surface, so the second head is not a third attribute.
- https://raw.githubusercontent.com/jwohlwend/boltz/main/docs/training.md recorded 2026-09-10
"To run training, you will need to download a few pre-processed datasets", covering both structures and MSAs, with configuration templates for the structure and confidence models.
- https://raw.githubusercontent.com/jwohlwend/boltz/main/README.md recorded 2026-09-10
Inference takes a YAML specification of the biomolecules with an optional MSA server; the Training section documents retraining, and the affinity outputs are predicted values rather than generated molecules.
Verified 2026-09-10