OpenFold3-Preview (github.com)

The AlQuraishi Lab and the OpenFold consortium have published OpenFold3-preview, an open-source, Apache‑2.0 research preview that aims to be a bitwise reproduction of DeepMind’s AlphaFold3. The release provides tools for fast, kernel‑accelerated inference and pipelines for predicting structures of proteins, RNA, DNA and small molecules, supporting monomers, complexes and protein‑ligand systems. Key technical features include MSA generation via ColabFold or JackHMMER/hhblits following the AlphaFold3 protocol, template support for protein monomers, cuEquivariance and DeepSpeed4Science kernel acceleration, multi‑query distributed predictions across GPUs, and memory‑constrained settings for smaller hardware. The package is installable via pip and documented on ReadTheDocs with HuggingFace examples to help developers get started quickly. Significance: OpenFold3-preview brings AlphaFold3‑level functionality into a fully open ecosystem, enabling researchers and companies to reproduce, extend, and integrate state‑of‑the‑art structure prediction into workflows without closed-source constraints. Preliminary benchmarks report competitive performance across CASP16, FoldBench and other datasets, and — uniquely among open models — parity with AlphaFold3 on monomeric RNA. The project solicits community feedback, welcomes contributions, and plans full parity across modalities, public training documentation, dataset release, and custom non‑PDB training workflows in future releases.