Protenix
High-accuracy open-source AlphaFold3 reproduction from ByteDance
What's new
Protenix-v2 (2026) is a 464M-parameter model with large gains on antibody–antigen complexes and training-free geometric guidance during diffusion.
Protenix is a production-grade, open reproduction of AlphaFold 3 built by ByteDance. Protenix-v1 was the first fully open-source model to outperform AlphaFold 3 across diverse benchmarks under the same data cutoff and inference budget.
Highlights
- Apache-2.0 code and weights with support for MSA, RNA MSA and templates.
- Protenix-v2 adds a training-free guidance module that enforces steric, torsion and bond constraints during diffusion sampling.
- Inference-time scaling (more samples) yields consistent gains on hard antibody–antigen targets.
Related tools
Free for academia
AI Models
AlphaFold 3
Predict the joint structure of proteins, nucleic acids and ligands
Web appModel weightsCommand line
Open source
AI Models
Boltz-2
Open-source structure and binding-affinity prediction, MIT licensed
Python libraryCommand lineModel weights
Open source
AI Models
OpenFold3
A community, bitwise-faithful open reproduction of AlphaFold 3
Python libraryCommand lineModel weights