Tasks and inputs
RFD4-Proteína is listed for Any-to-any, Data & control. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
RFD4-Proteína is an open Baker Lab and NVIDIA flow-matching model that designs protein sequence and all-atom structure together, under BSD-3-Clause.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Protein design pipelines have long split the work: generate a backbone, then guess a sequence that folds into it. RFD4-Proteína, built by the Baker Lab with NVIDIA's generative AI research group, designs both at once. It is a flow-matching model that outputs residue identities and all-atom coordinates together, conditioned on whatever context you need preserved: a binding target, an enzyme motif, a ligand or chosen structural features.
The bundle is a matched pair, a flow model plus the autoencoder whose latent space it generates in. Mixing a flow model with an autoencoder from another release does not throw an error; it produces nonsense. Keep them together.
| Release of 19 Oct 2026 | Previous release | |
|---|---|---|
| Residue vocabulary | 147 classes | 33 classes |
| Latent dimension | 32 | 16 |
| Molecule coverage | Proteins, nucleic acids, small molecules | Proteins |
| Stored parameters | 435,622,465 | 434,348,737 |
The current flow model is a pretrained base with a rank-32 LoRA merged in; both pieces also ship separately. Weights are FP32 EMA, and model.safetensors is 1.74 GB.
Proteina.from_pretrained from the RFD4 Python package, on Linux x86_64 with Python 3.12 and a CUDA 13-compatible driver.The licence is BSD 3-Clause for model and weights, one of the most permissive in this space. Optional third-party assets carry their own terms.
Pick a design task you have already solved in the lab or with an older tool, say a binder against a known target. Generate a few hundred candidates, filter them with your usual structure predictor and interface metrics, and compare hit rates with your current pipeline. Be cautious with nucleic-acid and small-molecule contexts. The unified vocabulary is new in this release, and that is where surprises will show.
For another open structure model, see Microsoft's BioEmu-2. NVIDIA's wider catalogue is on its provider page.
Compare the model on the work you need to complete.
RFD4-Proteína is listed for Any-to-any, Data & control. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under BSD-3-Clause. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Microsoft
BioEmu-2 is Microsoft Research's open ~500M diffusion model that samples ensembles of all-atom biomolecular structures, including complexes and ligands.