Nvidia and UW Built a Protein Design AI That Produces Six Times More Usable Designs Per Hour

Key points

• RFD4-Proteina produced 76.40 qualifying protein designs per GPU-hour at high thinking, versus 13.24 for the next-best method, across 24 protein-binding targets.
• Nvidia and the University of Washington's Institute for Protein Design co-developed the model as part of Nvidia's BioNeMo open-model programme.
• Free access on dedicated Nvidia RTX PRO 6000 Blackwell GPUs, the company's most powerful desktop graphics chips, runs from 21 October to 4 November 2026 via Hugging Face.
• After 4 November, a free shared-capacity version remains available on Hugging Face with no deadline.
• The model runs in a browser with no installation, making it accessible to researchers without dedicated AI hardware.

Scientists designing proteins for drug targets have always faced the same frustration: one tool shapes the protein's backbone, another strings together amino acids, a third checks atomic fit. Data bleeds away at every handoff.

RFD4-Proteina skips those handoffs entirely. One model generates the protein's sequence (its chain of chemical building blocks) and its full three-dimensional atomic structure at the same time, wrapping both steps into a single process.

The benchmark numbers are striking. Across 24 protein-binding tasks, RFD4-Proteina at its highest refinement setting produced 76.40 structurally distinct, passing designs per GPU-hour. The next-best method, Proteina-Complexa, managed 13.24. That's not a marginal improvement; it's roughly six times the output for the same compute cost.

How does the thinking setting work?

The model borrows a trick from large language AI: a "thinking" dial. Set it low and it generates fast, rough drafts. Turn it to medium or high and it refines each design as it goes, raising the chance a finished structure will pass the computational quality checks researchers use before moving to a lab. High thinking costs more processing time per design but delivers far more usable candidates per hour overall.

A separate option called reward-guided search goes further: the model generates several candidate proteins in parallel, scores them against user-defined properties like binding energy or hydrogen-bond count, and steers the next generation toward the winners. It's the model running its own internal audition before handing anything to the scientist.

Can anyone actually try it?

Yes, without installing a thing. Two versions live on Hugging Face, the open AI-model hosting platform: a dedicated Space running on Nvidia's RTX PRO 6000 Blackwell GPUs and a shared ZeroGPU Space that uses pooled capacity.

The dedicated Space is free until 4 November 2026. After that, the shared Space stays free indefinitely, subject to standard Hugging Face account limits.

Researchers wanting to go deeper can run the model in Python on Linux with a compatible Nvidia GPU. One caution worth flagging: generation is stochastic (it involves controlled randomness), so keeping the model version, seed value and hardware configuration is the only way to reproduce a promising result later.

AI2Day first covered protein design on 11 September 2026, and it's a beat where Nvidia keeps showing up. The retrieval models the company released the same day as RFD4-Proteina cut AI agent computing costs by 14 percent. The pattern holds: Nvidia isn't just selling chips; it's shipping the models that run on them.

For drug discovery teams, the practical upshot is a faster path from design goal to testable candidate at a compute cost that fits inside existing lab budgets. That matters most to smaller groups who can't afford to run a separate pipeline for every design stage. Whether these computational structures hold up in a real lab is still open. Every generated design needs experimental validation before anyone claims a working drug.