Remove the bucket validation link, source revision and historical/reproduction details from the optional-runtime paragraph. Keep a short description of supported runtimes and separate weights. All containers and other repository files remain unchanged.
Trim AF3 runtime extras while preserving tested CUDA 12/13 containers
Keep updated SIF, SquashFS and Docker images and their checksums at existing paths. Remove redundant OCI exports, build-recipe copies and JSON reports from the current runtime tree; validation remains linked from the public bucket. No container binaries, weights or AF2 files are changed.
Publish tested CUDA 12/13 AF3 and OpenFold3 runtimes with index-gap support
Both CUDA builds contain AF3 fork revision 8d36011. Validation: 32 source/index-feature cases and 48 AF3/OF3 GPU refolds. Model checkpoints, weights, and AF2 artifacts are unchanged.
The row listed ZeroGPU first, so the first Studio a visitor met was the one that spends their own Hugging Face quota and queues behind other users. The dedicated pool is the better default; ZeroGPU stays as the alternative for anyone who prefers it. The website made the same change in baker-laboratory/RFD4-Proteina-dev#771.
Cards: point links at rfd4-proteina.org docs and the new wordmark
RFD4-Proteina was merged into Diffusers upstream as huggingface/diffusers#15088 on 2026-10-20, so the pipeline no longer comes from a staging branch. It is not on PyPI yet, which the next Diffusers release fixes, so the quickstart installs from git and shows the plain pip line commented out beside it.
Checked rather than assumed, in a clean 3.12 environment: the git install imports RFD4ProteinaPipeline, and nvidia/RFD4-Proteina-Diffusers loads through it with all four components and no newly-initialized warnings. accelerate is listed because the load fails without it: "`low_cpu_mem_usage` cannot be False when `keep_in_fp32_modules` is True".
docs: link the ZeroGPU and dedicated-GPU Studio Spaces
The snippet is followed by "Partial diffusion and inverse folding work in this port", but its required list left coords_nm in conditioning_features, and the pipeline reads it as a named argument. Both of those tasks raise "partial and inverse generation require clean coords_nm" without it. A full run, which is what the snippet shows, ignores coords_nm, so the snippet behaves identically and the two tasks it advertises now work.
Model card: focus on native and Diffusers, drop the Transformers section
A single-chain 100-residue length specification, the same input the quickstart generates from. It carries no coordinates: the 100-100 expandable segment states the length and the model generates sequence and structure. Added for the Transformers integration's slow test, which fetches examples/monomer.cif from this repository and skips without it.
Host the refolding environment; drop the source snapshot
Adds openfold3/runtime/of3_af3_refolding_env.sif, the Apptainer environment for the OpenFold3 refolding oracle. Its SHA-256 (100679cc...) is the digest scripts/download_example_data.sh already pins, but the bucket URL that script curls now returns 404, so the documented setup path is broken; hosting it here restores it. Removes source/, since the code lives in the project repository and model repositories conventionally link to it rather than vendoring a tarball.
Publish inference-only checkpoints: drop trainer state (-4.53 GB)
The released -ema.ckpt files carried three copies of the weights' worth of data: the EMA state_dict, a duplicate SimpleEMA shadow copy, and optimizer state. Since generate.py treats -ema.ckpt as already-EMA, the shadow copy is redundant for every published use. Dropping callbacks, optimizer_states, lr_schedulers and loops takes the flow checkpoint from 4.21 to 1.74 GB and the autoencoder from 3.10 to 1.03 GB. All 1257 and 699 weight tensors are bit-identical to the originals, and generation at seed 42 reproduces the full-checkpoint output exactly. Full-checkpoint digests stay recorded in release-checkpoints.json and provenance.json; resuming a training run needs those, which are not distributed.
Document the Transformers checkpoint and its current limits
Transformers generation is supported from the transformers/ subfolder. Partial diffusion and inverse folding are not: they call the autoencoder's encoder, and rfd4-proteina-prep does not yet emit incident_bond_type_counts or ground_truth_element. Ligand and atomized inputs stop with an error rather than conditioning on unaligned geometry. Diffusers remains on the previous release.
Describe the three from_pretrained entry points and where each lives
Replaces the stale nvidia/RFD4-Proteina-hf reference. States that the root holds the native weights, that the Transformers export would collide with them on filenames and so goes in a subfolder, and that Diffusers needs model_index.json at a repo root.
Correct the card: only the native path runs this release
The card advertised Transformers and Diffusers usage that does not work with these weights. Both integrations target the previous release's 33-class residue vocabulary and 16-dimensional latent and reject this bundle's 147 classes and 32 dimensions, and nvidia/RFD4-Proteina-hf does not exist. States the supported path and says plainly that nvidia/RFD4-Proteina-Diffusers still serves the previous release.
Pin the release revision in the card and the checkpoint manifest
Replaces the bundle and the stable checkpoint aliases with the hero/unified-AE pair (latent dim 16 -> 32, residue vocabulary 33 -> 147 classes). Removes the previous release's dated filenames, which duplicated the aliases byte for byte; they remain available through the repository history.
27 days ago
Polish RFD4-Proteina card, examples and provenance
