RFD4-Proteina-Diffusers

RFD4-Proteina-Diffusers

Runs the RFD4-Proteina protein design model through the Diffusers library, so no native RFD4 package is needed. It suits researchers who design molecules around a target on a CUDA GPU.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository is the Diffusers version of RFD4-Proteina, a flow-matching model that designs protein sequences and all-atom structures together. It holds the same weights as the original release. You load it with one pipeline class, and the native RFD4 package is not needed at run time.

Who it suits: It suits computational biologists who already work with Diffusers and want to design molecules, such as a binder for a target. A CUDA GPU is recommended.

Industry | Who uses it | How they use it | Example
Biotech | Protein designers | Generate a binder around a target structure | Designing a binder of 50 to 120 residues for a given target
Research | Machine learning researchers | Run the model through a familiar library | Loading the pipeline in a Diffusers workflow
Structural biology | Structure analysts | Read designed coordinates back from CIF files | Inspecting the output structure of a design run

People who prefer the native RFD4 package can use the original release instead.

Where to use it:

• Python 3.11 or newer with Diffusers
• A computer with a CUDA GPU
• The rfd4-proteina-prep package, which prepares the inputs

Use case | Who | What they get back
Binder design | Protein designers | Designed atom positions and residue types, saved to a CIF file
Pipeline research | Machine learning researchers | A pipeline made of a transformer, an autoencoder and a scheduler

How to set it up:

• Python 3.11 or newer
• Diffusers installed from GitHub, since the card says the next release is still pending
• The matching rfd4-proteina-prep package, which supplies chirality and element features
• A CUDA GPU, which the card recommends

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install Diffusers from GitHub and the prep package, as the card's quickstart shows.

pip install -U huggingface_hub

hf download nvidia/RFD4-Proteina-Diffusers

How to use it: Load the pipeline in BF16 and move it to the GPU. Fetch an annotated CIF file that marks the parts to keep and the parts to design, then prepare its features. An older Diffusers fails to load this repository instead of sampling wrongly.

import torch
from huggingface_hub import hf_hub_download
from diffusers import RFD4ProteinaPipeline
from rfd4_proteina_prep import prepare_cif, write_cif

pipe = RFD4ProteinaPipeline.from_pretrained("nvidia/RFD4-Proteina-Diffusers", dtype=torch.bfloat16).to("cuda:0")

input_cif = hf_hub_download("nvidia/RFD4-Proteina-Diffusers", "examples/protein_binder/input.cif")
features, atoms = prepare_cif(input_cif, seed=42, device=pipe.device)

Tool | What it does
RFD4ProteinaPipeline | The pipeline that bundles the transformer, the autoencoder and the scheduler
prepare_cif | Turns an annotated CIF file into features for the pipeline
write_cif | Writes the designed result back to an annotated CIF file

Example: You fetch the example input from this repository, which sets a PD-L1 target and asks for a binder of 50 to 120 residues. You prepare its features and run the pipeline. The card's quickstart is cut off before the output step.

For newcomers and experts: This is a research pipeline, so start with the card's example input rather than your own molecule. Make sure the GPU, Python version and both packages are in place first.

The weights come from the pretrained base after a LoRA fine-tune was merged in. The original base and the adapter are kept on a separate branch. The released sampler recipe is called thinking_medium. It uses 60 steps and 180 trunk passes. The model runs in BF16, though the files are stored in FP32.

Annotated CIF structure
prepare_cif builds the features
Pipeline runs transformer and autoencoder
Result is written back to CIF

Good to know:

• License: BSD 3-Clause, covering the model and weights.
• Cost to run: none named. You need a suitable GPU, and the card names no paid service.
• Data it reaches: it downloads files from the Hugging Face Hub, and fetches the example input from this repository.
• Not for: older Diffusers versions, which cannot load this repository.

Title: Released sampler recipe
Steps: 60 steps
Trunk passes: 180 passes

On Hugging Face, 13 November 2026

• Library: diffusers
• Parameters: 178M
• Weights: about 714 MB in F32; about 89 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 403
• Likes: 5
• License: bsd-3-clause
• Languages: en
• Based on: nvidia/RFD4-Proteina
• Tags: diffusers, safetensors, biology, protein-design, biomolecular-design, structure-generation, flow-matching, annotated-cif

Links

• Model: https://huggingface.co/nvidia/RFD4-Proteina-Diffusers
• Files: https://huggingface.co/nvidia/RFD4-Proteina-Diffusers/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/RFD4-Proteina-Diffusers

Then follow the usage in its model card. Nothing in this kit runs by itself.