NVIDIA announces open RFD4-Proteina model debut

IPD and NVIDIA co-developed RFD4-Proteina as an open generative model.[2] NVIDIA introduced it as part of healthcare and life sciences collaborations at GTC Berlin on October 21, 2026.[1] The model combines programmable protein design with search capabilities in NVIDIA’s BioNeMo.[3] UW Medicine says the name brings together IPD’s RosettaFold Diffusion line and NVIDIA’s Proteina-Complexa models.[2] For teams assessing the release, the key distinction is that the announcement describes both design generation and ways to search and refine candidate designs.[3]

RFD4-Proteina generates protein sequences and all-atom structures together, conditioned on molecular context.[1] Researchers can use molecular inputs, constraints and scoring objectives to generate targeted designs, then search and refine them.[3] Its latent space includes standard and modified amino acids, sugars, DNA and RNA, and is supervised with force-field energies.[4] Recurrent refinement allocates computation adaptively, while constraints such as energetic or geometric criteria can guide search.[4] The NVIDIA NIM deployment supports seven tasks: generation, motif scaffolding, enzyme design, small-molecule binder design, antibody design, partial diffusion and inverse folding.[5]

Across computational benchmarks, UW Medicine reports design success rates five to ten times higher and candidate generation five to ten times faster than previous models.[2] NVIDIA also reports higher computational design success rates and throughput than evaluated baselines across protein-binding, small-molecule-binding, enzyme-scaffolding and DNA-binding benchmark suites.[3] On the protein-binding benchmark, the model achieved a 10.6% clustered pass rate versus 5.9% for the next-best model, at 42 times the throughput.[4] On one NVIDIA H200, a 250-token design takes five seconds in BF16 or 2.8 seconds after regional compilation.[1] These figures describe computational results, not a general guarantee of experimental performance.[2]

UW Medicine says RFD4-Proteina is immediately available as an open-source resource under BSD-3 terms, and NVIDIA lists it for commercial or non-commercial use.[2][5] The NVIDIA NIM deployment listing is dated October 20, 2026, at GTC Berlin.[5] UW Medicine says lab bench tests validated that the model’s protein outputs acted as anticipated in an experimental setting.[2] The project website reports that 12 of 18 designs bound below 1 micromolar and showed no detectable binding without calcium.[4] Those experimental findings provide a separate point of reference from the computational benchmark comparisons.[2][4]