jfold converted models

Released structure prediction models converted to jfold (JAX / Equinox), with each model's upstream sampler settings attached.

from jfold import pretrained

model = pretrained.load("boltz2")  # downloads boltz2/ from this repository on first use

Each model is a directory <name>/ with two files, written by jfold.io.save:

• weights.safetensors: every array, keyed by its path in the model (readable with numpy alone).
• model.json: the format version, the model's structure (class names, fields, plain values, array shapes and dtypes) and provenance metadata: the upstream source, the SHA-256 of the converted checkpoint, the upstream package version and the conversion date.

The ESMFold2 models share their language models, stored once each (esmc_6b, esmc_300m, esmc_600m) and attached by pretrained.load; the experimental type uses them rounded to bfloat16, as upstream runs it. The scaling-sweep checkpoints are its final step (1500k), whose ESMCs are the final ESMC weights. AtlasFold and AtlasFold-M share AtlasLM-3B the same way (atlaslm_3b).

Every conversion is checked against the upstream PyTorch model by jfold's checkpoint tests.

Models

Name	Upstream	Source	License	Size
boltz2	Boltz-2	boltz2_conf.ckpt (boltz)	MIT	2.0 GB
protenix_tiny	Protenix	protenix_tiny_default_v0.5.0	Apache-2.0	0.4 GB
protenix_mini	Protenix	protenix_mini_default_v0.5.0	Apache-2.0	0.5 GB
protenix_base	Protenix	protenix_base_default_v1.0.0	Apache-2.0	1.5 GB
protenix_base_20250630	Protenix	protenix_base_20250630_v1.0.0	Apache-2.0	1.5 GB
protenix_v2	Protenix	protenix-v2	Apache-2.0	1.9 GB
mol1_base	Mol-1	molence/Mol-1 mol-1-base	Apache-2.0	2.3 GB
opendde_v1	OpenDDE	opendde.pt	Apache-2.0	2.6 GB
opendde_abag	OpenDDE (antibody-antigen)	opendde_abag.pt	Apache-2.0	2.6 GB
openfold3	OpenFold3	of3-p2-155k.pt, EMA weights	Apache-2.0	1.5 GB
promera	ProMERA	bjing-mit/promera promera_2606.ckpt	MIT	1.8 GB
bioemu2	BioEmu2	microsoft/bioemu2 bioemu2-preview	MIT	2.1 GB
esmfold2	ESMFold2	biohub/ESMFold2	MIT	0.9 GB
esmfold2_fast	ESMFold2	biohub/ESMFold2-Fast	MIT	0.8 GB
esmfold2_experimental	ESMFold2 (experimental)	biohub/ESMFold2-Experimental	MIT	0.9 GB
esmfold2_experimental_fast	ESMFold2 (experimental)	biohub/ESMFold2-Experimental-Fast	MIT	0.7 GB
esmfold2_experimental_cutoff2025	ESMFold2 (experimental)	biohub/ESMFold2-Experimental-Cutoff2025	MIT	0.9 GB
esmfold2_experimental_fast_cutoff2025	ESMFold2 (experimental)	biohub/ESMFold2-Experimental-Fast-Cutoff2025	MIT	0.7 GB
esmfold2_experimental_fast_base300m	ESMFold2 scaling sweep, ESMC-300M	biohub/ESMFold2-Experimental-Fast-base300M-step1500k	MIT	0.7 GB
esmfold2_experimental_fast_base600m	ESMFold2 scaling sweep, ESMC-600M	biohub/ESMFold2-Experimental-Fast-base600M-step1500k	MIT	0.7 GB
esmfold2_experimental_fast_base6b	ESMFold2 scaling sweep, ESMC-6B	biohub/ESMFold2-Experimental-Fast-base6B-step1500k	MIT	0.7 GB
esmfold2_experimental_fast_base300m_confidence	ESMFold2 scaling sweep, ESMC-300M, with a confidence head (FastPLMs)	biohub/ESMFold2-Experimental-Fast-base300M-step1500k + the head of Synthyra/ESMFold2-300	MIT	0.7 GB
esmfold2_experimental_fast_base600m_confidence	ESMFold2 scaling sweep, ESMC-600M, with a confidence head (FastPLMs)	biohub/ESMFold2-Experimental-Fast-base600M-step1500k + the head of Synthyra/ESMFold2-600	MIT	0.7 GB
esmc_6b	ESMC-6B (float32)	the ESMC in biohub/ESMFold2	MIT	25.4 GB
esmc_300m	ESMC-300M (bfloat16 values)	the ESMC of the base300M scaling checkpoint	MIT	~1.3 GB
esmc_600m	ESMC-600M (bfloat16 values)	the ESMC of the base600M scaling checkpoint	MIT	~2.3 GB
esmc_300m_mlm	ESMC-300M with its masked-LM head (float32)	Biohub/ESMC-300M	MIT	1.3 GB
esmc_600m_mlm	ESMC-600M with its masked-LM head (float32)	Biohub/ESMC-600M	MIT	2.3 GB
esmc_6b_mlm	ESMC-6B's masked-LM head; its backbone is esmc_6b	Biohub/ESMC-6B	MIT	27 MB
atlasfold	AtlasFold (monomers)	SeonghwanSeo/atlasfold-260703	MIT	0.9 GB
atlasfold_m	AtlasFold-M (complexes)	SeonghwanSeo/atlasfold-m-260725	MIT	0.9 GB
atlaslm_3b	AtlasLM-3B (bfloat16 values, float32)	SeonghwanSeo/atlaslm-3b-base	MIT	12.2 GB

The weights keep their upstream licenses. Each <name>/ carries its upstream LICENSE and a NOTICE stating the conversion (arrays re-laid out and re-saved; values unchanged).