Boltz-2 (boltz2_conf and boltz2_aff checkpoints)
Release in the Boltz-2 family · version boltz-community/boltz-2: boltz2_conf.ckpt + boltz2_aff.ckpt
Maintained by Boltz (Boltz PBC)1235
The Boltz-2 weights published on June 6, 2025: a structure and confidence checkpoint (boltz2_conf.ckpt) and a separate binding-affinity checkpoint (boltz2_aff.ckpt), released together in one Hugging Face repository and loaded by the boltz package, which runs the latest model by default. The two checkpoints are assessed together because the affinity prediction runs as part of the same Boltz-2 pipeline.15310
- Model hub: Boltz-2 checkpoints (Hugging Face) (external site: huggingface.co)
- Repository: GitHub repository (jwohlwend/boltz) (external site: github.com)
- License: LICENSE (MIT) (external site: github.com)
- Paper: Boltz-2 technical report (bioRxiv, doi:10.1101/2025.06.14.659707) (external site: doi.org)
Availability and license
Overall availability
The boltz package downloads both checkpoints from a boltz.bio gateway and falls back to the ungated boltz-community Hugging Face repository. The optional automatic MSA step queries the public ColabFold server by default. No biosecurity-specific access conditions were found in the README or announcement.51310
Availability is separate from permission: read the license before using or redistributing.
The repository LICENSE is the MIT License, copyright 2024 by three named individual developers. The README states that all code and weights are provided under MIT for academic and commercial use, and the Hugging Face repository is tagged MIT. The June 2025 announcement says the model, weights, and training pipeline are released under MIT. Boltz PBC's hosted-service terms, which bar using outputs to build competing models, apply to its platform and API, not to this MIT release.4311013
Component reuse rights
- weights
- Reviewed qualifying license recorded — check scope and conditions
- code
- Reviewed qualifying license recorded — check scope and conditions
- data
- Unknown — no complete fact-level rights review
- documentation
- Unknown — no complete fact-level rights review
No complete system-rights review is recorded for this release.
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Both checkpoints are in an ungated Hugging Face repository tagged MIT.15 |
| Inference codeIs code for running the model published? | Public | The boltz package (PyPI and GitHub) provides the boltz predict command.39 |
| Training codeIs the code used to train the model published? | Partial | The repository includes a training script and configuration templates, but the provided structure configuration targets the Boltz-1 model class, and updated Boltz-2 training code and information are marked as coming soon.367 |
| Training-data informationDoes the information cover provenance, scope, acquisition, selection, labeling, processing, and where data or alternatives can be obtained? Access alone does not establish completeness. | Partial | The announcement mentions large collections of synthetic and molecular-dynamics data; the repository's data documentation covers Boltz-1 sources (PDB and OpenFold structures and MSAs). This confirms partial disclosure, not completeness.106 |
| Training-data accessCan the training data be obtained? This is independent of information completeness and reuse rights; original unshareable data need not be downloadable. | Unknown | Pre-processed Boltz-1 training data is downloadable per the training documentation; no download of the Boltz-2 training data was found.6 |
| Complete training pipelineIs the complete base-training and preprocessing pipeline published, including configuration? Fine-tuning code or an inference SDK alone is insufficient. | Unknown | The announcement says the training pipeline is released, but the repository documents only Boltz-1 data preprocessing and training; a complete Boltz-2 pipeline was not found.106 |
| Legacy data assessment (v0.1)Historical assessment combining download access and disclosure. Preserved for traceability; excluded from the v0.2 tier calculation. See the new separate assessments above. | Unknown | Not assessed. |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The announcement describes the added affinity module and the use of synthetic and molecular-dynamics training data at a high level; the technical report was not reviewed in full for this record.10 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | Boltz-2 results are published as plots in the README and evaluation documentation, but Boltz-2 evaluation files and scripts are marked as coming soon.38 |
What it is useful for
Predicting complex structures from YAML inputs and, for protein-ligand pairs, a binder probability for hit discovery and an affinity value for comparing related binders during hit-to-lead and lead optimization.3
Run and use notes
- The README installs Boltz with pip install boltz[cuda] and runs predictions with boltz predict; on CPU-only or non-CUDA hardware the [cuda] extra is dropped, and the README notes the CPU version is significantly slower. PyPI metadata requires Python 3.10 to 3.12; the latest release at review was 2.2.1 (September 8, 2025).39
- On recent NVIDIA GPUs Boltz uses NVIDIA cuEquivariance kernels, per the README.3
Organization context
Provenance and derivatives
Builds on Boltz-1, adding an affinity module and training on additional synthetic and molecular-dynamics data, per the Boltz-2 announcement. The technical report lists authors at MIT CSAIL, MIT Jameel Clinic, Recursion, Valence Labs, and ETH Zurich.10112
- Derived from: Boltz-1 (external site: huggingface.co) — Earlier open model from the same team; no separate catalog record.
Other releases in the Boltz-2 family
No other releases in this family have been assessed.
What this catalog does not know
- Training-data access: unknown.
- Complete training pipeline: unknown.
- Legacy data assessment (v0.1): unknown.
Have a primary source? How to report a correction.
U.S. eligibility
Eligible · basis: U.S.-governed project
Developed by the Boltz team at MIT Jameel Clinic with Recursion and now presented by Boltz PBC, whose terms give a Massachusetts notice address, as one of its open-source models. The technical report also lists authors at Valence Labs and ETH Zurich. The repository sits under an individual developer's GitHub account and its license names three individual copyright holders.101213411
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.