sokrypton/af3-any-model

AlphaFold3-family model weights, converted to run in one package

Seven published AlphaFold3-architecture models, converted from their original PyTorch checkpoints into the AlphaFold 3 JAX/Haiku parameter format so that a single codebase runs all of them from the same input JSON.

Code: sokrypton/alphafold3, branch af3-any-model

# the weights are fetched on first use; nothing to download by hand
python run_alphafold.py --model=boltz2 --json_path=fold_input.json \
                        --output_dir=out --norun_data_pipeline

Each model ships two files: <model>.bin.zst, the parameters, and <model>.shapes.json, the parameter tree derived from the graph plus a record of which converter produced the blob and when — it is what makes loading skip jax.eval_shape, and it names anything a conversion did not cover. All seven conversions cover every parameter the graph asks for.

What is here, and whose it is

These are DERIVED works: the same trained parameters, rewritten into another framework's layout. Each remains under its original licence and belongs to its original authors.

file	model	authors	licence	original
openfold3.*	OpenFold3	AlQuraishi Lab / OpenFold Consortium	Apache-2.0	openfold3
intellifold2.*	IntelliFold-v2	IntelligenAI	Apache-2.0	intelligenAI/intellifold
opendde.*	OpenDDE	Aureka Research	Apache-2.0	aurekaresearch/OpenDDE
boltz2.*	Boltz-2	Wohlwend et al., MIT	MIT	jwohlwend/boltz
protenix2.*	Protenix-v2	ByteDance	Apache-2.0	bytedance/Protenix
rosettafold3.*	RoseTTAFold3	Institute for Protein Design, UW	BSD-3-Clause	RosettaCommons foundry
chai1.*	chai-1	Chai Discovery	Apache-2.0	chaidiscovery/chai-lab

If you use one of these, cite the model's own authors.

AlphaFold 3's own parameters are not here and will not be. Google DeepMind requires you to request them directly; point --model_dir at your own copy.

Notes on two of them

• chai-1 also needs std_conformers.npz (in this repo, fetched with it) and ESM2 token embeddings, which are most of its token feature stream. Without the embeddings it is a different model — see converters/esm_embed.py and --esm_embeddings.
• protenix2 was converted from a community mirror of ByteDance's release, since the official CDN was unreachable; the SHA256 was verified against the CDN copy while it still resolved.

How they were made

python -m converters.convert --model NAME --out DIR in the repo above, which downloads the published checkpoint, converts it, and writes the shape manifest. The conversion is not mechanical — residue alphabets differ between codebases, as do the row/column conventions of pair projections, and getting either wrong is silent. OF3_AF3_PORTING_NOTES.md and docs/ported_models.md record what each one required.

Sanity check

All seven fold 6MRR (a de novo designed 68-residue protein) from a single sequence, scored against the crystal structure, at 3 recycles and 1 sample:

model	CA-RMSD	pLDDT		model	CA-RMSD	pLDDT
boltz2	0.52 Å	96.8		opendde	1.59 Å	92.0
protenix2	0.67 Å	84.8		intellifold2	1.63 Å	85.5
rosettafold3	0.99 Å	81.5		openfold3	1.74 Å	78.6
				chai1	1.75 Å	84.5

(AlphaFold 3 itself gets 0.61 Å on the same input, for reference.)