For two years, matching AlphaFold3's structure-prediction quality meant either waiting for a restricted release or accepting a visible accuracy gap. Boltz closed it with an MIT-licensed model anyone can run commercially, then pushed past pure structure into the question drug hunters actually care about: how tightly does this molecule bind?
What Sets It Apart
- Boltz-1 was the first fully open-source model to approach AlphaFold3 accuracy on biomolecular complexes, jointly predicting structures of proteins, nucleic acids, and small-molecule ligands. Unlike AlphaFold3's restricted weights, it ships under the MIT license for both academic and commercial use.
- Boltz-2 adds an affinity module that predicts protein-ligand binding affinity at an accuracy approaching free-energy perturbation (FEP) simulations while running over 1000x faster, turning a multi-hour atomistic calculation into a screenable prediction.
- The engineering is built for scale, not just demos: batched YAML inputs, NVIDIA cuEquivariance kernels for GPU acceleration, and even Tenstorrent hardware support, so the same model spans a single complex or a high-throughput virtual screen.
Who It's For
Great fit if you are a computational biologist or drug-discovery team that needs AlphaFold3-class predictions you can deploy, fine-tune, and ship in a commercial pipeline without licensing friction. Look elsewhere if you need experimentally confirmed structures: these are predictions, and the affinity estimates still trade some accuracy for the 1000x speedup, so wet-lab validation stays essential for anything consequential.
