NVIDIA Nemotron 3.5 Embed 8B
Create multilingual embeddings for semantic search across document text, page images and combined image-text content.
RFD4-Proteína
huggingface.co
Generate protein sequences and all-atom structures from molecular design constraints.
Visit website
Pricing
Free
Available on
Linux, Python, NVIDIA GPU
Category
Education & Studies
What is RFD4-Proteína?
RFD4-Proteína is a downloadable biomolecular design model that jointly generates protein sequences and all-atom structures. It accepts annotated CIF structures or YAML conditions and returns CIF designs.
For protein engineers and computational biologists, examples cover unconditional monomer generation, protein binders and enzyme motif scaffolding.
Key features
Joint sequence and structure generation
Generate residue identities and atomic coordinates together.
Annotated molecular constraints
Mark structural regions to preserve or design.
Multiple design tasks
Start from examples for protein binders, enzyme scaffolds or unconditional monomers.
Search guided by molecular evaluations
Guide inference-time search with geometric and energy evaluations, while recurrent refinement allocates additional computation.
Who is it for?
Protein engineers exploring binders
Generate candidate binders around a specified target for subsequent evaluation.
Enzyme researchers scaffolding motifs
Explore protein designs around a catalytic motif.
Computational biology developers
Integrate constrained generation into a candidate assessment pipeline.
How to use RFD4-Proteína
- Prepare a Python 3.12 environment and supported NVIDIA GPU using the selected runtime's CUDA instructions.
- Install the native package, or install Diffusers from GitHub until its next release includes the required pipeline.
- Prepare an annotated CIF or YAML condition specification.
- Load matched checkpoints, generate candidates and inspect the CIF outputs.
- Evaluate selected candidates structurally and experimentally.
RFD4-Proteína pricing
Free
Downloadable weights use BSD 3-Clause; compute costs are separate. NIM deployment follows separate terms.
Check current pricing ↗What to know before you use it
- NIM runs each design request on one GPU. Additional GPUs serve concurrent requests.
- Flow-model and autoencoder releases must match; mixing them can silently invalidate results.
- Optional refolding and scoring require separate oracle environments and weights, with their own license terms.
Frequently asked questions
Can it design around an existing motif?
Yes. The model card provides an enzyme motif scaffolding example.
Which installation route should I choose?
The native package supports Python workflows. Diffusers currently requires installation from GitHub until the next release.