Boltz-1

Released in November 2024 by MIT's Jameel Clinic for Machine Learning in Health (preprint bioRxiv 2024.11.19.624167), Boltz-1 is the first fully open-source model to match AlphaFold3-level accuracy on biomolecular structure prediction — proteins, RNA, DNA, ligands, and glycans. Unlike AlphaFold3, whose weights carry academic-use restrictions, Boltz-1 ships code and weights under the MIT license with a pip-installable package, and its optimized diffusion architecture predicts a complex in roughly 30–60 seconds.

The Boltz team has since commercialized follow-ups (Boltz-2 adds binding-affinity prediction; BoltzGen adds therapeutic design) through a public-benefit corporation, but Boltz-1 itself remains freely usable commercially. Honest limitations: like all diffusion-based folders it can emit physically impossible artifacts on very large complexes, and very large structures are constrained by its cropping window — predicted structures still need experimental or physics-based validation before drug-design decisions.