v0.6.2

[0.6.2] - 2026-08-07

OpenFold3 lands: tt-bio predict --model openfold3 folds proteins, RNA and DNA with the
OpenFold Consortium's AlphaFold3 reproduction, with per-chain MSAs and optional per-chain
templates, on the same scheduler, multi-card fan-out and MSA cache as Protenix-v2. Polymer
chains only — ligands, covalent bonds and --write_pae raise or are declined rather than
silently degrading. OpenFold3 is the one model whose weights tt-bio does not download: the
consortium's checkpoint is a public release you fetch yourself — point OF3_CKPT at it or
drop it at ~/.boltz/of3-p2-155k.pt.

Added

• OpenFold3 — tt-bio predict --model openfold3 folds proteins, RNA and DNA with the
  OpenFold Consortium's AlphaFold3 reproduction, with per-chain MSAs and optional per-chain
  templates. It rides the same scheduler, worker, multi-card fan-out and MSA cache as
  Protenix-v2. Parity-gated against the official CPU reference on seven legs (see
  docs/implementation-parity.md); the on-device diffusion sampler runs in fp32 by default
  to match the reference's own boundary. Polymer chains only — ligands, covalent bonds and
  --write_pae raise or are declined rather than silently degrading. This is the one model
  whose weights tt-bio does not download: set OF3_CKPT. (9c10f08f9)

Fixed

• OpenFold3's vendored data pipeline imports pydantic, pdbeccdutils, func_timeout,
  networkx and packaging at module load, and none were declared, so --model openfold3
  failed on a clean pip install tt-bio. The vendored Apache-2.0 license text now ships in
  the wheel too.
• tt-bio predict --model openfold3 showed no trunk or diffusion phase on the live
  progress view: the worker's progress adapter was never passed to the model. Wired
  through the trunk and sampler loops, mirroring Protenix-v2 (df0ed79f).

Performance

• Protenix-v2 and OpenDDE at ~300 residues: the 298-aa GPU scaling gap is closed, 1.18x on
  the landed code (ec28f3d2; measured 62934f2f). At that size the Pairformer trunk is
  74.5% of step time and scales as N² (46c4fe29); the SDPA chunk gate, the trimul
  in0_block_w and the transition chunk were re-tuned for it (7349f407, ec50003d).

Gates and documentation

• The README, predict --help, the parity docs and the OpenFold3 port doc now match what the
  shipped model actually does, including template support (a6403435, a35f7a3b,
  1327a080, 2b0671ee, 92d5b58a, 50bc0bee).
• kabsch_rmsd was labelled a CA RMSD but is computed over every atom name; the label is
  corrected at the source and on the OpenFold3 rows. The other models' rows follow in a
  deliberate pass — the affinity pocket legs use a different, genuinely CA-based scorer.
• OpenFold3 is enrolled in the accuracy, perf and UX release gates (86932b20, 1f236666,
  2a20f028).
• scripts/fetch_parity_fixtures.sh verifies the fetched tarball by its hash field instead
  of the sidecar's recorded absolute path, which exists only on the machine that generated
  it (f9d0afc1).

Release gate (Blackhole P150a on tt-quietbox)

Host suite: 225 passed / 23 skipped / 1 xfailed with 25 failures triaged one file per
process — 24 are environment or harness artefacts, not code regressions (qb1's boltzgen
transformers env gap, the parent-holds-device false-failure class, a stale July OF3 dev golden
that the P8+ tests skip without, and a 4e-6 PCC wobble on the confidence leg). The 25th was
real: ec50003d's transition big-chunk gate admitted the Protenix-v2 N=512 pair shape
(W=512), which overflowed in-block L1 (test_fold_512_no_oom). Fixed as e6678e21 (gate
tightened to W<=384, keeping the 298-aa fast path); the test passes on the fixed tree.
Packaging guard: 16/16 data files and 36/36 declared dependencies in the wheel and sdist.

Accuracy gate — every shipped fold architecture folded end-to-end with production sampling
(200 steps, 5 samples) and checked against a per-model ground-truth floor, not
self-consistency:

model	RMSD (A)	TM	floor	result
boltz2	1.373	0.945	<=3.0 / >=0.75	PASS
esmfold2	4.462	0.563	<=8.0 / >=0.40	PASS
esmfold2-fast	1.769	0.910	<=4.5 / >=0.60	PASS
protenix-v2	2.459	0.799	<=6.0 / >=0.50	PASS
opendde	1.352	0.955	<=6.0 / >=0.50	PASS
openfold3	2.042	0.845	<=3.5 / >=0.70	PASS

Parity gate (scripts/full_parity_gate.py, 30 legs): 24 PASS, 6 GAP, 0 DRIFT, 0 ERROR,
every GAP reproducing a committed GAP-evidenced record. All seven OpenFold3 legs pass on the
release tree (ubq X=1.34 A within R=1.64/D=0.80; 8hel-msa and 9bk6 included). Four records are
newly evidenced this release: protenix-prot-msa, opendde-prot-prod and opendde-trpcage-nomsa
(c97076c0, root-caused to ba6ede96's intended AttentionPairBias unfusing, device samples
inside each reference's own inter-seed spread and every numerator inside the committed noise
floor), and openfold3-7xi5-notmpl (af8f886d, fold accuracy verified against RCSB 7XI5
directly: all five device seeds at 0.589-0.609 A aligned CA-RMSD, the CPU reference's own
spread being 0.422-0.895 A).

Performance gate: every model within +/-15% of its committed baseline on both cards.
OpenFold3 vs its committed p300c baseline on tt-quietbox2: 2.191 structures/s vs 2.142
(+2.3%) — the 298-aa shared kernel-gate work does not regress it. Full p150a leg on
tt-quietbox: 15/15 PASS, worst |delta| 9.3% (boltzgen), everything else within +/-4%.
OpenFold3's p150a entry (0.990 structures/s) is a first seeded baseline, disclosed as such
(cdebd298) — it is compared against itself this release; the p300c comparison above is
this release's regression evidence for the model.

UX gate: PASS on the tag tree — every model's live progress advances through trunk
and diffusion, outputs parse, the CLI surface behaves. It earned its keep: OpenFold3's
first hardware UX leg caught the missing progress wiring (df0ed79f), re-run green after
the fix.