NVIDIA opens RFD4-Proteina protein design model on Hugging Face Spaces, free dedicated GPUs until Nov 4

NVIDIA is sponsoring free access to a dedicated Hugging Face Space for RFD4-Proteina from October 21 to November 4, 2026. The Space runs on NVIDIA RTX PRO 6000 GPUs. RFD4-Proteina is a BioNeMo open model for protein design, co-developed by NVIDIA and the University of Washington's Institute for Protein Design (IPD).

What the model does

RFD4-Proteina combines protein sequence and all-atom structure generation in a single model. This replaces the usual pipeline of separate backbone, sequence, and all-atom design models. According to NVIDIA, it can generate proteins in contexts that include other proteins, small molecules, nucleic acids, glycans, metal ions, and modified amino acids such as a phosphorylated residue.

The architecture pairs a flow-matching model with an autoencoder:

• The autoencoder represents each amino acid (standard or modified), sugar, DNA nucleotide, or RNA nucleotide as an anchor-atom coordinate (the alpha carbon for amino acids) plus a learned latent vector encoding chemical identity and local geometry.
• The flow model generates anchor coordinates and latent vectors together.
• A decoder expands them into all-atom 3D structures and sequences.

Parameter count, training data, and licensing terms are not stated in the source material.

Thinking setting

The model adds a "thinking" setting (low, medium as default, high) that controls how much each design is refined during generation. More refinement raises the likelihood of designs passing in silico filters but increases compute per design. A separate, optional reward-guided search scores designs with external evaluators, such as a structure prediction model or a physics-based energy calculator. It can target interface area, hydrogen bonding, or ligand binding energy.

Benchmark claim

In a design benchmark across 24 protein-binding targets, NVIDIA reports that RFD4-Proteina at high thinking produced 76.40 structurally distinct designs per GPU-hour meeting its computational criteria. Proteina-Complexa, the second-strongest method in the benchmark, produced 13.24. The criteria, based on a structure prediction refold, are:

• binder pTM ≥ 0.80
• minimum interface PAE ≤ 1.5 Å
• target-aligned binder backbone RMSD ≤ 2.5 Å

These are NVIDIA's figures, and the results are computational. The company notes that experimental validation is needed to confirm binding.

Access options

Option	Hardware	Cost
Dedicated RFD4-Proteina Space	NVIDIA RTX PRO 6000	Free Oct 21 to Nov 4, 2026; pricing after that not yet disclosed
ZeroGPU Space	Shared capacity	Free, subject to Hugging Face account quotas

Both Spaces run AtomWorks Studio, a browser-based viewer and annotation editor for molecular inputs. Users start from a supplied example, such as the built-in PD-L1 binder task. They define a target, a binding region, and hotspot residues, then generate designs. Outputs download as annotated CIF files that carry the design context. NVIDIA also points users to a protein design competition sponsored by Anthropic and Adaptyv.

For programmatic use, the documented runtime requires Linux x86_64, Python 3.12, a supported NVIDIA GPU, and a CUDA 13-compatible driver. A CUDA 12 build is also provided, pairing the all-cuda12 extra with a cu126 PyTorch. NVIDIA recommends recording the input, model and software revisions, seed, sampling settings, and hardware for each run, because generation is stochastic.

What this means

The notable change is architectural. Generating sequence and all-atom structure jointly, with explicit conditioning on ligands, nucleic acids, glycans, and modified residues, removes the handoff losses in multi-model design pipelines. The thinking setting brings the test-time compute tradeoff familiar from LLMs into structural biology, with a clear cost-for-quality dial.

The headline benchmark is a vendor-run comparison on 24 targets, scored by in silico filters rather than wet-lab results. The 5.8x per-GPU-hour gap over Proteina-Complexa needs independent replication, and computational pass rates do not guarantee binding. The free two-week dedicated Space and the permanently free ZeroGPU option lower the barrier to that kind of testing, particularly with the Anthropic and Adaptyv competition running alongside.