Can an open-source model truly match AlphaFold3’s performance? ByteDance says yes. Their new Protenix-v1 model, released under Apache 2.0, achieves AF3-level accuracy across proteins, DNA, RNA, and ligands — while keeping everything open for research and production use. This isn’t just another AlphaFold clone. Protenix-v1 includes a complete training pipeline, pre-trained weights, and a browser-based server for interactive predictions. The real differentiator? A rigorous evaluation toolkit called PXMeter that benchmarks over 6,000 complexes with transparent metrics. ## Why It Matters AlphaFold3 revolutionized biomolecular structure prediction but remained largely closed. Protenix-v1 democratizes this capability: - Full open stack: Code, weights, training pipelines — all available on [GitHubhttps://github.com/bytedance/Protenix){rel=“nofollow”} - Fair comparisons: Model matches AF3’s training data cutoff (2021-09-30) and inference budget - Extensible: Designed for customization, not just inference The research team claims Protenix-v1 is the first open-source model to outperform AlphaFold3 on diverse benchmark sets under matched constraints. ## The Technical Core Protenix-v1 implements an AF3-style diffusion architecture for all-atom complexes: - Parameters: 368M (matching AF3’s undisclosed scale class) - Coverage: Proteins, nucleic acids, ligands - Inference scaling: Log-linear accuracy gains with more sampled candidates The included PXMeter v1.0.0 toolkit provides: - Curated benchmark dataset (6,000+ complexes) - Time-split and domain-specific subsets - Unified metrics: complex LDDT, DockQ ## Beyond Structure Prediction The Protenix ecosystem extends beyond prediction: - PXDesign: Binder design suite with 20–73% experimental hit rates - Protenix-Dock: Classical docking framework - Protenix-Mini: Lightweight variants for cost-effective inference ## Key Takeaways 1. AF3-class, fully open: First open-source model matching AlphaFold3 performance 2. Fair benchmarking: PXMeter enables transparent, reproducible evaluations 3. Production-ready: Includes training code, weights, and a web server 4. Extensible ecosystem: Covers prediction, docking, and design The model is available at [protenix-server.comhttps://protenix-server.com/login){rel=“nofollow”}, with the full stack on [GitHubhttps://github.com/bytedance/Protenix){rel=“nofollow”}.