A new Open Data release via Registry of Open Data on AWS makes OpenFold3 training datasets publicly available to support reproducible research, benchmarking, and retraining for new questions.

Industry adoption is highlighted, with SandboxAQ noting previous integration of OpenFold advances into their AQAffinity model and continued use of OpenFold3 for expanded capabilities.

OpenFold3 remains an open-source deep learning system for cofolding, capable of predicting 3D structures of biomolecular complexes from sequence and inputs, including protein–small molecule and protein–nucleic acid interactions.

Antibody–antigen complex prediction is identified as a challenging frontier, with plans in 2026 to expand data, benchmarks, and model improvements focused on immune-relevant complexes.

OpenFold3 is positioned as competitive with AlphaFold3 across multiple modalities and tasks, as reflected in updated benchmarks in the accompanying white paper.

OpenFold Consortium announced a major OpenFold3 update and the public release of training data and full-stack tooling for reproducible biomolecular AI, announced in Berkeley on March 16, 2026.

The project reiterates its mission to keep foundational biology AI open, reproducible, auditable, and accessible to academia, biotech, pharma, and nonprofit research, supported by ongoing compute, datasets, and software engineering investment.

A new OpenFold3 portal provides onboarding, deployment guidance, reference pipelines, evaluation scripts, dataset access, and a public community channel for support and issue triage.

The release bundles full-stack resources—training datasets, model weights, training and inference code, evaluation scripts, and permissive licenses—to enable reproduction and community extension.