And a global consortium of industry and academic partners
The OpenFold Consortium has announced the release of OpenFold3, the most extensively trained open-source co-folding model ever created, and a major leap forward for democratizing in silico biomolecular structure prediction.
Built by a global network of researchers, and with SandboxAQ as a key contributor, OpenFold3 accurately predicts the 3D structures of proteins and their interactions with other molecules, from small-molecule drugs, antibodies, other proteins, or nucleic acids, enabling faster, more cost-effective, and more precise discovery of new therapeutics.
OpenFold 3 brings state-of-the-art, open-source co-folding AI to every researcher worldwide, matching the performance of the Nobel-recognized AlphaFold 3 while remaining fully open and accessible.
40 Million
Synthetic structures and 300K publicly available, experimental structures
40 Million
CPU hours for data generation
5 Million
GPU hours beyond model training
See Biology
In-silico
Accelerate Molecular Design
Discover Better Medicines
Drug discovery depends on understanding how a protein binds to other molecules. Proteins are large sequences of amino acids folded in a three-dimensional space. The 3D folded structure of the protein dictates its functionality, and the co-folded structure of a protein and a candidate drug molecule determines whether a potential treatment will succeed or fail.
Until now, obtaining these structures experimentally has been slow, expensive, and complex. OpenFold3 changes that.
OpenFold3 is a co‑folding model that predicts the structure of a protein bound to another molecule, such as a small‑molecule drug, an antibody, another protein, or a nucleic acid. These structures are essential to design effective, specific therapies.
As both a core member and key technical contributor of the OpenFold Consortium, SandboxAQ has been instrumental in the conception, development, and optimization of OpenFold3.
Built at the intersection of AI modeling, computational chemistry, and NVIDIA DGX infrastructure, our team engineered robust molecular AI systems designed for openness and scale. We conducted large-scale model training, systematic ablation studies, and performance optimization to push the limits of predictive accuracy. Every component, from code to compute, is designed for reproducibility, transparency, and impact across the global scientific community.
For more information:
OpenFold3 is now available fully open-sourced under the permissive Apache 2.0 license on GitHub and Hugging Face for academic and industry researchers and developers globally. A GPU-optimized, containerized and streamlined inference microservice is available as a NIM through the Nvidia BioNemo platform for scalable workflows.
Explore OpenFold3 through:
with Nvidia and the OpenFold Consortium
