BoltzGen weights (tt-bio mirror)

Mirror of the BoltzGen binder-design weights (plus the Boltz-2 folding/affinity checkpoints the design pipeline depends on), packaged for use with tt-bio on Tenstorrent hardware. The files are byte-for-byte identical to the upstream releases; this repo hosts them on the Hugging Face Hub so tt-bio can fetch them with huggingface_hub like every other model it runs.

Files

File	Description
boltzgen1_diverse.ckpt	BoltzGen design model — diverse protocol
boltzgen1_adherence.ckpt	BoltzGen design model — adherence protocol
boltzgen1_ifold.ckpt	BoltzGen inverse-folding (sequence design) model
boltz2_conf_final.ckpt	Boltz-2 structure model used for the folding step
boltz2_aff.ckpt	Boltz-2 affinity model used for the affinity step
mols.zip	CCD molecule / component library

Note: boltz2_aff.ckpt here is the checkpoint shipped with the BoltzGen pipeline and differs from the one in moritztng/boltz-2.

Usage

tt-bio downloads these automatically when you run a design:

tt-bio gen run examples/design.yaml --output out/

Or fetch a single file directly:

from huggingface_hub import hf_hub_download
path = hf_hub_download("moritztng/boltzgen", "boltzgen1_diverse.ckpt")

Credit & license

BoltzGen is developed by Hannes Stärk and the BoltzGen authors and released under the MIT License (https://github.com/HannesStark/boltzgen). The bundled boltz2_conf_final.ckpt and boltz2_aff.ckpt are Boltz-2 weights by Jeremy Wohlwend, Gabriele Corso, Saro Passaro and the Boltz authors, also MIT (https://github.com/jwohlwend/boltz). All weights and the accompanying LICENSE are redistributed here unmodified under those terms. Please cite BoltzGen and Boltz if you use these weights.