Copyright 2026 Anthropic, PBC

This tree contains 36 model packages and one shared library (common/opt_core).

Original code in this tree — each package's opt/ directory, its run.sh
command-line wrapper, configuration, container recipes, documentation and
tests, the shared library common/opt_core (including its kit_template/ and
tools/ directories) and common/mps_packing — is licensed under the Apache
License, Version 2.0 (the "License"); see the LICENSE file in this directory.
You may not use these files except in compliance with the License. Unless
required by applicable law or agreed to in writing, software distributed under
the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
CONDITIONS OF ANY KIND, either express or implied. Nothing in this NOTICE
relicenses third-party code or data.

UPSTREAM RELEASES UNDER stock/

Each package's stock/ directory carries the upstream project at a pinned
release, under that project's own licence: the release archive or wheel as
published (apart from the exceptions listed below) and, in most packages, an
unpacked reading copy of it or of the upstream repository at the pinned commit
(stock/src/). The package's LICENSE file names the upstream licence, whose
text travels inside stock/ and under the package's third_party_licenses/.
Beside the upstream files every stock/ directory holds two files written for
this tree, PINS.json (versions and SHA-256 pins) and check_pins.py; some also
hold this tree's install helpers and patch files
(af3_jax/stock/unpack_src.sh and stock/patches/,
af3_torch/stock/fetch_upstream.py and stock/patches/,
openfold3_ob0/stock/install_upstream.py), the PyPI metadata and provenance
records of the pinned wheel (opendde/stock/pypi_*.json), the
model card and configuration of the pinned checkpoint (enformer/stock/snapshot/,
from the model repository named in enformer/STOCK.md) and, in e1, the Hugging
Face hub kernel that stock E1 loads (e1/stock/hub_kernel/: the kernel's files
and their BSD-3-Clause LICENSE, plus SOURCE.md written for this tree).

The upstream files under stock/ are byte-identical to the pinned releases with
these exceptions, each declared in the package's STOCK.md:
  - af3_jax: stock/src/src/alphafold3/model/network/template_modules.py
    carries the package's one declared bug-fix patch (nine added lines); the
    archive stock/alphafold3-bc32b22f.tar.gz is unmodified.
  - pxdesign: the PXDesign reading copy and archive leave out two font files
    of upstream's pxdesign/pxd_server/ directory (Helvetica-Regular.ttf,
    TimesNewRoman.ttf: proprietary typefaces of Linotype and Monotype that
    upstream bundles for its server page), which nothing in the package
    reads (pxdesign/STOCK.md).
  - colabdesign: the BindCraft reading copy and archive leave out the two
    pre-compiled third-party executables upstream bundles under functions/:
    dssp, which run.sh install fetches from the BindCraft repository at the
    pinned commit and verifies by digest, and DAlphaBall.gcc, which is not
    distributed, not fetched and not used by any route of the package
    (colabdesign/STOCK.md, colabdesign/THIRD_PARTY_NOTICES.md).
  - complexa: the Proteina-Complexa archive leaves out one member,
    src/proteinfoundation/result_analysis/sc, a pre-compiled third-party
    executable that upstream's README says is obtained separately; nothing
    the package runs reads it (complexa/STOCK.md 'Stock exceptions').
PXDesignBench is not carried in this tree; pxdesign's run.sh install fetches it
from upstream at its pinned commit, under its own licence.

THIRD-PARTY MATERIAL OUTSIDE stock/

Third-party source also lives inside the packages' own directories. Every such
file is listed — with its upstream path, licence and copyright line — in the
package's licenses/THIRD_PARTY_NOTICES.md (common/opt_core/THIRD_PARTY_NOTICES.md for
the shared library), and its licence text is under the package's
third_party_licenses/ or beside the file. The classes and their locations:

(a) Vendored third-party code, unmodified:
  - chrombpnet/opt/kit/torch/vendor/: bpnet-lite 1.0.0 and tangermeme 1.4.1
    (MIT) with their dist-info licence files;
    chrombpnet/opt/kit/torch/chrombpnet_k1/_vendor/typing_extensions.py
    (PSF-2.0);
  - caliby/opt/forward/xattempt_addon/stock/: eight files of Caliby
    (Apache-2.0), Protpardelle-1c (MIT) and the Chroma modules Caliby carries
    (Apache-2.0, Generate Biomedicines), kept as the unmodified counterparts of
    the modified copies in ../fast/;
  - common/opt_core: trimmed portions of FlashAttention-3's hopper/utils.h
    and hopper/softmax.h (two copies of fa3_utils.h, BSD-3-Clause notice
    kept) inside the triangle-attention kernel sources.

(b) Modified copies of upstream files (the upstream header kept; the changes
    marked in the file or described beside it):
  - af2ig/opt/forward/af2ig_kit/patches/patched_files/ (dl_binder_design,
    MIT, and the AlphaFold 2 files it carries, Apache-2.0, DeepMind);
  - af3_jax/opt/forward/fast_inference/patches/patched_files/run_alphafold.py
    (AlphaFold 3, Apache-2.0, DeepMind, as forked by alphafold3-open);
  - af3_torch/opt/forward/af3t/af3_torch/xfold/ (the xfold model files, part
    of them modified; see the note on their headers below);
  - borzoi/opt/datapath/pipeline_tf/v17/borzoi_sad.py (Borzoi, Apache-2.0,
    Calico);
  - caliby/opt/forward/xattempt_addon/fast/ (the eight files named in (a));
  - protenix_v1/opt/forward/v05_addon/lib/detpatch/scatter_utils.py and
    protenix_v2/opt/forward/flashpairformer/src/detref/scatter_utils.py
    (Protenix, Apache-2.0, ByteDance);
  - rfdiffusion3/opt/forward/xattempt_addon/patched/rfd3/ and
    rosettafold3/opt/forward/rf3_xattempt_addon/patched/rf3/ (foundry,
    BSD-3-Clause);
  - esmfold2/opt/forward/fast_inference/upstream/ (unified diffs against the
    carried transformers fork; not applied by the package).

(c) Prebuilt GPU binaries compiled from third-party or third-party-derived
    source, each beside its source or a README/SOURCE file that names it:
  - the fused LayerNorm extension carried by protenix_v1
    (opt/forward/v05_addon/lib/fastln_prebuilt/), protenix_v2
    (opt/forward/flashpairformer/third_party/fastln_prebuilt*/) and opendde
    (opt/forward/fast_inference/levers/LNSTREAM/prebuilt/): Protenix's
    layer-norm CUDA sources (Apache-2.0, ByteDance; portions by OneFlow and
    HPC-AI Technology, Apache-2.0), themselves modified from NVIDIA Apex
    (BSD-3-Clause), with this tree's changes; the modified sources sit in the
    patched_src/ or src/ directory beside each binary;
  - chrombpnet/opt/kit/torch/triton_cache_*/: Triton compiler output for the
    package's own kernels, including Triton's runtime helper cuda_utils.so
    (Triton, MIT);
  - the shared library's extensions, cubins and XLA custom-call libraries under
    common/opt_core/opt_core/kernels/ and the packages' own .so/.cubin files
    are compiled from this tree's sources; the NVIDIA CUTLASS / CuTe,
    FlashAttention-3, CCCL, PyTorch, pybind11, jaxlib XLA FFI and TensorFlow
    headers they were built against (some binaries therefore contain object
    code instantiated from those headers' templates), the Triton compiler that
    produced the cubins, and the CUDA runtime code nvcc links in (statically in
    four XLA libraries of the shared library and in protenix_v2's
    libtriatt_procuda_sm90a.so) are named with their licences, binary by
    binary, in the same notices files.

(d) Files of this tree that restate, fuse or re-express statements of the
    upstream model code they accelerate (kernels, levers, drivers) and
    therefore derive from it. Such files exist in every package and in
    common/opt_core; they derive from the project the package wraps and in
    places from further projects (AlphaFold 2 and 3, OpenFold3, Protenix,
    FlashAttention, PyTorch, xFormers, CPython), and each is attributed in
    the package's licenses/THIRD_PARTY_NOTICES.md with the upstream file it derives
    from. Several Protenix-derived directories in protenix_v1, protenix_v2,
    pxdesign and opendde, and the OpenFold3-derived directories under
    openfold3/opt/forward/ and openfold3_ob0/opt/forward/, carry a copy of the
    upstream Apache-2.0 LICENSE (with ByteDance's or the AlQuraishi
    Laboratory's copyright line) for that reason; the files in them are this
    tree's work except where the notices name an upstream file.

Notes on particular components:
  - af3_torch: the model files carried from xfold keep the CC BY-NC-SA 4.0
    header text they inherited from AlphaFold 3's initial source release;
    xfold's own licence, like the licence of the AlphaFold 3 source carried in
    af3_jax, is the Apache License, Version 2.0
    (af3_torch/THIRD_PARTY_NOTICES.md lists the files and marks the ones this
    tree modified).
  - esmfold2: the carried stock/transformers-ef32577f.tar.gz is a snapshot of
    a fork of Hugging Face transformers whose repository is no longer public;
    it is redistributed under its Apache-2.0 licence, which travels inside the
    archive (esmfold2/STOCK.md, esmfold2/THIRD_PARTY_NOTICES.md).
  - No file of NVIDIA cuEquivariance (source, binary or packaged data) is
    distributed in this tree; NVIDIA libraries are installed from NVIDIA's own
    channels at install or build time under NVIDIA's terms. The tuning tables
    in cuEquivariance's cache-file format under protenix_v2 were generated by
    running that kit's own workloads through the library's tuner, and the
    kernels in common/opt_core that return results bitwise identical to a
    cuEquivariance operation were written independently against that
    operation's observable inputs and outputs.
  - Apart from the CUDA runtime code linked into the prebuilt binaries named
    in (c), NVIDIA's CUDA toolkit, cuDNN, NCCL, TransformerEngine and CUDA
    base images are not carried in this tree; the install routes fetch them
    from NVIDIA's own distribution channels under NVIDIA's terms.

PROFLUENT-E1 (package e1)

e1 builds on Profluent-E1 (Profluent Bio Inc.), which is distributed under the
Profluent-E1 Clickthrough License Agreement together with the Profluent-E1
Attribution Guidelines; copies exactly as released are e1/stock/src/LICENSE,
e1/stock/src/NOTICE and e1/stock/src/ATTRIBUTION. Recipients of e1 receive
Profluent-E1 under that Agreement; downloading or using Profluent-E1 is
acceptance of it. The Agreement and Guidelines require attribution to
"Profluent-E1" in distributions and documentation and, for redistribution in
object form, a display of "Profluent-E1" each time the resulting program runs:
e1/README.md carries the statement "This project builds upon Profluent-E1 under
the terms of the Profluent-E1 Clickthrough License Agreement", and e1/run.sh
prints a one-line Profluent-E1 attribution naming the Agreement file at the
start of every command. The Profluent-E1 model weights are not carried; e1's
run.sh install --weights fetches them from the provider.

MODEL WEIGHTS

No package distributes the weights of the model it wraps, in this tree or in
images built from it; users obtain them from the original providers under
those providers' terms, and each package's STOCK.md names the source and the
terms. Small parameter files that upstream projects ship inside their own
source releases travel inside the carried archives exactly as released:
colabdesign (ColabDesign archive: eight ProteinMPNN parameter files, MIT, and
colabdesign/af/weights/template_dgram_head.npy, a small parameter array
ColabDesign ships for AlphaFold 2's template distogram head; AlphaFold 2
parameters are CC BY 4.0, DeepMind), mosaic (ProteinMPNN and AbMPNN weight
files, MIT and CC BY 4.0, and a CC BY 4.0 trigram table), borzoi (Baskerville
archive: tests/cpu/data/eval/model.h5, a small trained test model released with
Baskerville, Apache-2.0) and enformer (enformer-pytorch's
precomputed/tf_gammas.pt, a table of positional constants, MIT). They are
listed in those packages' licenses/THIRD_PARTY_NOTICES.md files.

DATA FILES IN CARRIED RELEASES AND IN THE PACKAGES

Several upstream releases ship example or test data drawn from public sequence
and structure databases; those files travel inside the carried archives and
reading copies exactly as released, under the databases' own terms:
  - rosettafold3: foundry archive member
    models/rf3/docs/examples/msas/3en2_A.a3m.gz, an alignment of UniRef100
    sequences (UniProt Consortium, CC BY 4.0); rfdiffusion3's foundry archive
    carries the rfd3 documentation's example input structures (wwPDB-derived).
  - pxdesign: stock/src/PXDesign/examples/ (also inside the archive): the PDL1
    example alignments pairing.a3m and non_pairing.a3m hold UniRef100,
    UniProtKB and UniParc sequences (CC BY 4.0), MGnify protein sequences
    (EMBL-EBI: CC0 1.0, releases before 2019 CC BY 4.0) and sequences of the
    ColabFold environmental database, which is built from BFD, MGnify and
    metagenome assemblies (CC BY-SA 4.0); 5o45.cif is a wwPDB entry.
  - openfold3, openfold3_ob0: stock/src/openfold3/tests/test_data/ holds
    upstream's test alignments for PDB entry 2Q2K (hits drawn from UniRef90 and
    UniProtKB, CC BY 4.0; MGnify, CC0 1.0 or CC BY 4.0 by release; BFD and
    Uniclust30, CC BY-SA 4.0; and template searches against PDB chain
    sequences), wwPDB mmCIF files and the NumPy/pickle snapshots produced by
    upstream's tests.
  - esm_if1: the fair-esm archive's examples/ hold alignments built from
    UniRef (CC BY 4.0) and from Uniclust30 (hhblits_uniclust_2017_10_*.a3m;
    Uniclust, CC BY-SA 4.0), the TEM-1 beta-lactamase mutational-scanning
    data of the variant-prediction example (BLAT_ECOLX_Ranganathan2015*.csv,
    P62593.fasta, as released with fair-esm), wwPDB coordinate files (4UV3,
    5YH2, 2N2U) and the lm-design example data.
  - af3_jax: the AlphaFold 3 archive's src/alphafold3/test_data/ holds
    miniature_databases/ — small excerpts that upstream includes for its
    tests, with the terms upstream states for them in
    af3_jax/stock/src/README.md: BFD (Steinegger and Söding, CC BY-SA 4.0;
    the excerpt as modified by Google DeepMind is stated there as CC BY 4.0),
    MGnify, PDB mmCIF and seqres, Rfam and RNAcentral (CC0 1.0), UniProt and
    UniRef90 (CC BY 4.0) and NCBI NT (NCBI's terms) — and upstream's
    featurised example and stored model outputs (*.pkl, *.json) generated
    from them.
  - esmc, esmfold2: the esm archive's esm/data/ holds entry, hierarchy and
    keyword tables derived from InterPro (EMBL-EBI, CC0 1.0) as released with
    esm.
  - complexa and genie3: the target structure sets of Proteina-Complexa
    (assets/target_data/) and of Genie's binder benchmark
    (data/design/binder_design/binderbench/) are crops, excerpts or prepared
    models of wwPDB entries and published de novo binder targets as released by
    those projects (Proteina-Complexa states CC BY 4.0 for its prepared data in
    licenses/license_datasets.txt inside the archive); af2ig's dl_binder_design
    archive carries its examples/ (wwPDB-derived target crops with the authors'
    designed binder models) and the AlphaFold 2 test structures it vendors;
    colabdesign's BindCraft copy carries example/PDL1.pdb (a crop of a
    deposited PD-L1 structure); chrombpnet's archive carries the motif tables
    of its data/ directory (position-weight matrices compiled by upstream from
    JASPAR, CC BY 4.0, and HOCOMOCO, with upstream's own bias motifs); borzoi's
    archives carry Borzoi's data/ tables and Baskerville's tests/cpu/data/
    (including hg38_1m.fa.gz, a one-megabase excerpt of the GRCh38 human
    reference assembly, and miniature training shards); opendde's reading copy
    carries examples/2lwu.cif and 7pzb.pdb (wwPDB entries); flashzoi's wheel
    carries Borzoi's track table (precomputed/targets.txt) and gpnstar's
    archive a 128-bp alignment fixture of the store named below.
wwPDB archive data are free of all copyright restrictions (CC0 1.0).

Test and example inputs written into the packages themselves (tests/inputs,
public_inputs, inputs_public, inputs/msa, inputs/targets, tests/specs,
tests/fixtures and examples directories) are public records or synthetic:
coordinate files are wwPDB entries or crops of them, named by their PDB
identifiers, with two exceptions — boltzgen's PD-L1 test specification
(opt/forward/fast_inference/tests/specs/PDL1_IgV_Q9NZQ7_18-132.pdb) is a
trimmed AlphaFold Protein Structure Database model (AF-Q9NZQ7-F1; CC BY 4.0,
DeepMind and EMBL-EBI), and rfdiffusion1's inputs_public/insulin_target.pdb is
a computed structure model (its header reads EXPDTA THEORETICAL MODEL) — the
insulin receptor target example distributed with RFdiffusion (BSD-3-Clause) —
not a wwPDB entry. Test sequences are those entries' sequences or synthetic
(e1's warm-up set); the alignment files (.a3m) in these directories are single-
sequence (query-only) inputs and contain no database sequences; flashzoi's
warm-up window (opt/forward/kits_v1_25/canary_window_0.npz) is a one-hot
524,288-bp window of the GRCh38 human reference assembly; gpnstar's
examples/msa/ is a 0.14-MiB excerpt of the public 100-vertebrate alignment
store described in gpnstar/examples/README.md (Hugging Face dataset
songlab/multiz100way, dataset card licence MIT; alignment origin UCSC hg38
multiz100way). Where a directory needs entry-level attribution it or its parent
carries a README, NOTICE, PROVENANCE or SOURCES file.

COPYLEFT AND SHARE-ALIKE TERMS INSIDE CARRIED RELEASES

Two kinds of file with copyleft or share-alike terms travel inside carried
releases exactly as their upstream projects ship them, unmodified.
OpenStructure's stereo_chemical_props.txt data table (LGPL-3.0), the file
AlphaFold 2 and OpenFold obtain from OpenStructure, travels in four byte-
identical copies: in colabfold's two wheels (colabfold-1.6.1, member
colabfold/openstructure/ with the LGPL text beside it;
alphafold_colabfold-2.3.13, member alphafold/common/), in af2ig's
dl_binder_design archive (member af2_initial_guess/alphafold/common/) and in
complexa's Proteina-Complexa archive (member
community_models/openfold/resources/); colabfold/THIRD_PARTY_NOTICES.md
describes the file and its licence. The alignment data drawn from BFD,
Uniclust30 or the ColabFold metagenomic sets named above (esm_if1's Uniclust30
alignment, the bfd_uniclust_hits.a3m files of openfold3 and openfold3_ob0, the
environmental-database entries of pxdesign's non_pairing.a3m) are under CC BY-
SA 4.0. No other GPL- family licence applies to a file in this tree or inside
the carried archives.
