Is this page helpful?

Release Notes for NVIDIA NIM for RFD4-Proteina#

Release 1.0.0#

Summary#

Initial release of the NVIDIA NIM for RFD4-Proteina, a generative protein design model developed by the Baker Laboratory in collaboration with NVIDIA.

Key Features#

  • Seven design tasks: generation, motif_scaffolding, enzyme_design, small_molecule_binder, antibody_design, partial_diffusion, and inverse_folding. Refer to Overview for what each does.

  • Two input dialects: a self-contained annotated mmCIF, or a condition spec plus its referenced structure files.

  • Three execution backends: oss_eager and oss_torch_compile (the upstream reference implementation, uncompiled or torch.compile-compiled), and bioir, an NVIDIA-accelerated trunk built on the BioNeMo Inference Runtime. bioir is the default; selectable per deployment via NIM_BACKEND. Refer to Configuration Alternatives.

  • Three named recycle schedules, spelled as the checkpoint’s own configuration: thinking_low (110 trunk passes), thinking_medium (185, the default) and thinking_high (295), all at 60 integration steps. A sampling.recycle_schedule block accepts custom [steps, depth] bands for deployments that want to explore beyond them, bounded by NIM_MAX_TRUNK_PASSES. Refer to Recycle Schedules.

  • Reward-guided beam search: an optional search block turns a request into a beam search over the generation trajectory, branching and pruning candidates against a weighted combination of metrics rather than returning one unguided sample. A search runs one replicate at a time (sampling.batch_size: 1); refer to Replicates and Batching.

  • Batched replicates: sampling.num_replicates is evaluated in a single forward pass by default, and sampling.batch_size trades that back for per-design seed replay or for a lower memory ceiling. The replicate count carries no fixed limit, since what a GPU can serve depends on the token count; an oversized request is answered with out_of_memory (HTTP 413) naming the fields to change. Refer to Replicates and Batching.

  • Design metrics selected by tag, including TMol-based interface energies (ligand_interface). A tag the running deployment cannot satisfy is rejected at request time rather than silently producing no output. The antibody tag is a refolding tag despite not being named *_refolding, and like the others needs a mounted oracle; refer to Oracles.

  • Mountable refolding oracle: an operator-mounted structure-prediction oracle can score each design and report eleven in-silico pass criteria (unconditional, PPI, antibody, DNA, motif-scaffold, internal-coordinates, small-molecule-binder, AME, AME-no-ligand, DISCO, developability) as part of the response’s aggregated metrics. No oracle is built in; refer to Oracles.

Model Checkpoints#

This release serves the following checkpoints.

Component

Checkpoint

Flow model

chk_epoch_00000169_step_000000340000-ema.ckpt

Autoencoder

2026_09_04_release_ae_z16_400k-ema.ckpt

Note

These checkpoints are not generally available. They are pre-release checkpoints, available only through this NIM, and are not part of the Baker Laboratory’s public RFD4-Proteina release. Results from this release can differ from the public model.

Every response reports both names in its reproducibility metadata.

Known Issues#

  • No refolding oracle is built in. Every *_refolding metric tag is rejected with 422 and all eleven in-silico pass criteria are unavailable until an operator mounts a reward.py; a reward-guided search can only weight the non-oracle metrics until then. This is a deliberate design choice rather than a defect – refer to Oracles – but it does mean in-silico success rate is not reportable out of the box.

  • The first request against each token-count bucket is slow. The NIM reports ready as soon as its weights load and does not pre-warm; the first request of a given size pays a compilation cost once per bucket per container start, measured between 1.9x and 9.8x steady state. To avoid it, send one request per bucket you expect to use before serving traffic. Refer to Cold-Start Behavior.

  • A reward-guided search cannot run batched replicates. search with sampling.num_replicates > 1 requires sampling.batch_size: 1, because a batch leaves no way to attribute a beam’s survivors to the replicate they came from. Refer to Replicates and Batching.

  • The public RFD4-Proteina checkpoints are not supported. The checkpoints in the Baker Laboratory’s public release (2026_10_15_*) use a new model architecture, and this release cannot load them. Support is planned for a later release. Refer to Using Your Own Model Checkpoints.

Roadmap#

A future release will support the Baker Laboratory’s public RFD4-Proteina checkpoints (2026_10_15_*), with their expanded residue vocabulary.