ByteDance drops Protenix-v1, an open-source model outperforming AlphaFold 3

Protenix-v1 is an open-source biomolecular structure prediction model from ByteDance that predicts complex 3D structures spanning proteins, nucleic acids, and small molecule ligands.

If you work in computational biology or drug discovery, AlphaFold 3 changed the game for multi-molecule prediction, but strict access controls left researchers hunting for fully reproducible open alternatives. Protenix-v1 fills that gap, outperforming AlphaFold 3 across diverse benchmarks under the exact same data cutoff and parameter scale.

What it brings to the table

Protenix-v1 runs on a base architecture of roughly 368M parameters, closely mirroring AlphaFold 3. It supports structural templates alongside protein and RNA multiple sequence alignments (MSAs).

For tricky targets like antigen-antibody complexes, the model leverages inference-time scaling. By dialing up your sampling budget from a few candidates to several hundred seeds, accuracy scales up in a steady log-linear trajectory.

ByteDance released two flavors of the v1 model. One uses the original September 2021 data cutoff for direct apples-to-apples academic comparison with AlphaFold 3, while an applied release updates training data through June 2025 for better real-world performance.

Alongside the model, the repo includes PXMeter, an open-source evaluation suite built with clean benchmarks that strip away experimental artifacts and non-biological noise.

How to run it and license terms

The entire Protenix-v1 package, including code and weights, is released under the permissive Apache 2.0 license for both research and commercial applications.

Crucially, ByteDance did not just dump model weights. The team open-sourced the complete training data pipeline and MSA pipeline, complete with a local ColabFold-compatible sequence search workflow.

You can pull it directly from PyPI via pip install --upgrade protenix. Running predictions requires only a formatted JSON input file, which you pass into the protenix pred CLI command alongside your output directory and chosen model tag.

Keep in mind that while Protenix-v1 is fully open, ByteDance's newer 464M Protenix-v2 remains proprietary with restricted weights, making v1 the primary workhorse for open deployment.

Why it matters

Ever since AlphaFold 3 landed, researchers have pushed for an open engine that replicates its multi-entity capabilities without cloud rate limits or closed data pipelines. Protenix-v1 delivers on that promise.

In our view, the real breakthrough is not merely squeaking past AlphaFold 3 on benchmark tables, but handing the community the full recipe from raw data processing to inference. If your team is designing binders, screening targets, or fine-tuning specialized biophysical models, having an Apache 2.0 pipeline of this caliber running on your own metal is something you should test immediately.