Folding proteins at home now requires a permission slip

What it does

The repository implements the full AlphaFold 3 inference pipeline for predicting biomolecular structures from sequence input. It separates the CPU-bound data pipeline—handling genetic and template searches—from the GPU inference stage, allowing the two phases to run on different hardware if needed. The pipeline will not function without model parameters received directly from Google under their separate terms of use.

The interesting bit

Unlike a typical open-source drop, the model weights are gated behind an application form that Google reviews at its sole discretion, turning local inference into a permissioned activity. The code itself is licensed under CC-BY-NC-SA 4.0, which explicitly excludes commercial use—a significant restriction for a tool aimed at biomolecular interaction prediction.

Key highlights

• Complete local inference pipeline for the AlphaFold 3 model described in the Nature paper on biomolecular interaction prediction
• Separates the CPU-bound data pipeline (genetic and template search) from GPU inference, so the stages can run independently
• Model weights are not bundled; access requires a direct request to Google and acceptance of separate terms of use
• A hosted non-commercial alternative exists at alphafoldserver.com with a reduced feature set for ligands and modifications
• Code is under CC-BY-NC-SA 4.0 and outputs carry an explicit “not for clinical use” disclaimer

Caveats

• Model parameters must be received directly from Google; you cannot simply download them from the repository
• The license and terms of use explicitly prohibit commercial use and clinical application
• Known issues are documented separately, so check the docs before assuming a bug is yours

Verdict

Worth exploring if you are an academic or non-commercial researcher who needs more control than the web server offers and does not mind waiting for Google to approve weight access. Skip it if you are looking for an unencumbered, commercially usable model or a turnkey download.