RFD4-Proteina: an open protein model
Between 19 and 21 October 2026 NVIDIA and the Baker lab at the University of Washington's Institute for Protein Design released RFD4-Proteina, an open model (BSD-3-Clause licence) that designs protein sequences and all-atom structures together, with code on GitHub and weights on Hugging Face. NVIDIA announced it at GTC Berlin on 21 October as available through nine partners, among them Hugging Face, Google Cloud, Databricks, Claude Science and OpenAI.
Why it matters
An open model for designing proteins and other biomolecules, with a paper that reports laboratory tests, lands at once in the agent and cloud tools scientists already use. What is reported is the authors' own comparison with other methods and a first demonstration of a designed enzyme; independent use has not yet been described.
What the pages say. NVIDIA's release: RFD4-Proteina is "a new NVIDIA BioNeMo open model" that combines programmable protein design with search; according to the "research draft" it achieved higher computational design success rates and generation throughput than the baselines evaluated, across four benchmark suites (protein binding, small-molecule binding, enzyme scaffolding, DNA binding); the page gives no numbers. Available "today" through Amazon Bio Discovery, Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio. The project page rfd4-proteina.org: the paper "Biomolecular design through latent reasoning with RFD4-Proteina" (Corley, Didi, Kazman, Brent and others; last author David Baker; a preprint with no date on the page); a shared 32-dimensional chemical and conformational latent space for amino acids, sugars, DNA and RNA; by the authors, on four in silico benchmarks the model has higher success rates at greater throughput than all methods evaluated; experimental validation spans binding, catalysis and "a first proof-of-concept demonstration of a de novo designed enzyme for active pharmaceutical ingredient synthesis". The Hugging Face repository nvidia/RFD4-Proteina: 435,622,465 stored parameters, 1.74 GB, BSD-3-Clause; the code RosettaCommons/RFD4-Proteina on GitHub is under the same licence. What the record does not claim. The benchmark numbers, the paper and the laboratory results were not checked: they are the authors' claims, and the paper's PDF was not read. The same NVIDIA release announced a collaboration with Siemens Healthineers on radiotherapy; there the models are only being "explored", and that part is not in this record.