OpenFold recorded 3,200 GitHub stars and 24 consortium partners as of December 2025, establishing itself as the premier open-source alternative to AlphaFold2. The AlQuraishi Lab at Columbia University developed this Apache 2.0-licensed implementation that achieves accuracy parity with DeepMind’s proprietary model while enabling commercial applications. Training time dropped from 8 days to 10 hours through infrastructure optimization, with market projections showing the AI protein folding sector expanding from $1.5 billion in 2024 to $15.3 billion by 2034.
OpenFold Key Statistics
- OpenFold maintains 3,200+ GitHub stars and 642 forks as of December 2025, with version 2.2.0 supporting PyTorch 2 and CUDA 12
- The OpenFold Consortium expanded to 24 partners in 2025, adding eight new members including Bristol Myers Squibb, Novo Nordisk, and Johnson & Johnson
- OpenFold achieves 68.6-78.8% mean GDT-TS accuracy on CASP15 benchmarks, matching AlphaFold2 performance on 50% of evaluated domains
- Training infrastructure optimizations reduced processing time from 8+ days on 128 A100 GPUs to 10 hours using 2,080 H100 GPUs
- OpenFold3 trained on 300,000+ experimental structures combined with 13 million synthetic structures, utilizing up to 256 GPUs continuously
OpenFold Accuracy and Performance Benchmarks
OpenFold demonstrated accuracy parity with AlphaFold2 in the May 2024 Nature Methods publication. Both models performed equivalently on exactly 50% of 90 CASP15 domain targets, with mean GDT-TS accuracy scores of 68.6-78.8% for OpenFold compared to 69.7-79.2% for AlphaFold2.
The evaluation used 10,000 bootstrap samples to establish 95% confidence intervals. OpenFold shows particular strength in predicting extremely long protein chains, handling sequences exceeding 4,000 residues on a single A100 GPU.
| Performance Metric | OpenFold | AlphaFold2 |
|---|---|---|
| Mean GDT-TS Accuracy (CASP15) | 68.6-78.8% | 69.7-79.2% |
| Performance Parity on Targets | 50% | 50% |
| CASP15 Domains Evaluated | 90 | 90 |
| Maximum Residue Prediction (Single A100) | 4,000+ | Limited |
OpenFold Training Infrastructure and Speed Improvements
NVIDIA optimizations transformed OpenFold training efficiency. The original AlphaFold2 training required 11+ days on 128 TPUs, while OpenFold initially needed 8+ days on 128 A100 GPUs.
Scaling to 2,080 H100 GPUs reduced training time to 10 hours. The ScaleFold implementation achieved a record 7.51 minutes in the MLPerf HPC v3.0 benchmark, representing more than 6x speedup over baseline configurations.
Researchers discovered that approximately 90% of model accuracy can be achieved in just 3% of the total training time, enabling rapid experimentation cycles.
| Training Configuration | Duration | Hardware |
|---|---|---|
| Original AlphaFold2 | 11+ days | 128 TPUs |
| OpenFold Baseline | 8+ days | 128 A100 GPUs |
| Optimized OpenFold | 10 hours | 2,080 H100 GPUs |
| MLPerf HPC v3.0 Record | 7.51 minutes | ScaleFold |
OpenFold Memory Efficiency and Computational Performance
OpenFold delivers substantial computational efficiency improvements through custom CUDA attention kernels and DeepSpeed integration. The DS4Sci_EvoformerAttention kernel reduces peak device memory requirements by 13x while providing 15% faster training and 4x speedup during inference.
Custom attention kernels modified from FastFold use 4x less GPU memory than equivalent FastFold implementations and 5x less than stock PyTorch versions. Inference speed reaches up to 2x faster compared to AlphaFold2.
OpenFold Consortium Growth and Industry Partners
The OpenFold Consortium welcomed eight new industry partners in April 2025, including Bristol Myers Squibb, Novo Nordisk, and Johnson & Johnson. The consortium now comprises 24 partners spanning academia, technology corporations, startups, and pharmaceutical companies.
Six global pharmaceutical firms participate in the consortium. The OpenFold3 development involved contributions from more than 40 institutions worldwide, demonstrating collaborative open-source structural biology research.
| Consortium Metric | Figure |
|---|---|
| Total Consortium Partners (2025) | 24+ |
| Global Pharma Companies | 6 |
| New Members Added (April 2025) | 8 |
| Contributing Institutions (OpenFold3) | 40+ |
| Founding Year | 2022 |
OpenFold3 Development and Training Dataset
Released in October 2025, OpenFold3 represents a significant advancement in biomolecular structure prediction capabilities. The model trained on over 300,000 publicly available experimentally determined structures combined with an OpenFold-curated synthetic database containing more than 13 million structures.
The training infrastructure utilized up to 256 GPUs operating continuously for several weeks. AWS infrastructure achieved 85% cost savings through strategic use of EC2 Capacity Blocks and Spot Instances, requiring 32 p5en.48xlarge EC2 instances.
| OpenFold3 Specification | Value |
|---|---|
| Experimental Structures Used | 300,000+ |
| Synthetic Structure Database | 13 million+ |
| Training GPUs Required | Up to 256 |
| EC2 Instances (p5en.48xlarge) | 32 |
| Training Duration | Several weeks (24/7) |
| AWS Cost Savings vs On-Demand | 85% |
| License | Apache 2.0 |
OpenFold GitHub Repository and Community Metrics
OpenFold maintains strong community engagement with over 3,200 GitHub stars and 642 forks as of December 2025. The repository publicly released approximately 400,000 multiple sequence alignments and PDB70 template hit files under a permissive license, hosted through the Registry of Open Data on AWS.
The latest v2.2.0 release introduced PyTorch 2 and CUDA 12 support, along with updated dependencies including numpy 2 and PyTorch Lightning 2.5. The framework operates on PyTorch 2 with CUDA 12 compatibility.
| GitHub Metric | Count |
|---|---|
| GitHub Stars | 3,200+ |
| GitHub Forks | 642+ |
| Training MSAs Released | 400,000+ |
| Current Version | v2.2.0 |
| Framework | PyTorch 2 / CUDA 12 |
OpenFold Licensing Terms vs AlphaFold3
OpenFold3 distinguishes itself through Apache 2.0 licensing for all software components, enabling both academic and commercial applications. AlphaFold3 restricts usage to limited academic research and prohibits commercial use.
OpenFold3 allows users to test, train with new datasets, develop applications, and adapt the model freely. The Linux Foundation classified OpenFold3 as a Class 1 open-source system, ensuring full transparency and reproducibility for the global research community.
| Feature | OpenFold3 | AlphaFold3 |
|---|---|---|
| License Type | Apache 2.0 | Restricted Academic |
| Commercial Use | Permitted | Not Licensed |
| Training Code Available | Yes | No |
| Model Modification | Permitted | Not Permitted |
| Linux Foundation Classification | Class 1 Open-Source | N/A |
AI Protein Folding Market Growth and Projections
The AI protein folding market recorded $1.5 billion in 2024 value and projects expansion to $15.3 billion by 2034 at a 26.1% compound annual growth rate. Protein structure prediction tools segment shows even stronger growth at 28.6% CAGR, expanding from $329 million to $4.07 billion over the same period.
North America dominates the protein engineering market with a 40.6% share in 2024, driven by robust biotechnology infrastructure and significant AI investment. The drug discovery and development segment holds 45% of the 3D protein structures analysis market.
| Market Segment | 2024 Value | Projected Value | CAGR |
|---|---|---|---|
| AI Protein Folding Market | $1.5 billion | $15.3 billion (2034) | 26.1% |
| Protein Structure Prediction Tools | $329 million | $4.07 billion (2034) | 28.6% |
| Protein Engineering Market | $3.6 billion | $8.2 billion (2033) | 9.5% |
| 3D Protein Structures Analysis | $2.8 billion | $6.88 billion (2034) | 9.4% |
OpenFold Research Impact and Citation Metrics
The OpenFold publication in Nature Methods (May 2024) became a foundational reference for trainable protein structure prediction systems. AlphaFold2, which OpenFold reproduces, accumulated nearly 43,000 citations as of November 2025.
Over two million researchers utilize AlphaFold-based systems for protein predictions across drug discovery, medical treatments, and basic research. OpenFold’s contribution to democratizing this technology enabled countless academic institutions and smaller biotechnology companies to access state-of-the-art structural prediction capabilities without proprietary restrictions.
The Federated OpenFold3 Initiative involves five pharmaceutical companies training on 4,000 to 8,000 protein-drug pairs from their proprietary libraries, representing a novel approach to improving model performance while maintaining data confidentiality.
FAQ
How many consortium partners does OpenFold have in 2025?
OpenFold has 24 consortium partners as of 2025, including six global pharmaceutical companies. The consortium added eight new members in April 2025, including Bristol Myers Squibb, Novo Nordisk, and Johnson & Johnson.
What is OpenFold’s accuracy compared to AlphaFold2?
OpenFold achieves 68.6-78.8% mean GDT-TS accuracy on CASP15 benchmarks, matching AlphaFold2’s 69.7-79.2% performance. Both models perform equivalently on 50% of 90 evaluated domain targets according to Nature Methods publication from May 2024.
How long does it take to train OpenFold?
OpenFold training time ranges from 8+ days on 128 A100 GPUs to 10 hours using 2,080 H100 GPUs. The MLPerf HPC v3.0 benchmark record achieved training in just 7.51 minutes with optimized ScaleFold implementation.
What license does OpenFold3 use?
OpenFold3 uses Apache 2.0 licensing for all software components, permitting both academic and commercial applications. This contrasts with AlphaFold3’s restricted academic license that prohibits commercial use and model modification.
How many structures was OpenFold3 trained on?
OpenFold3 trained on over 300,000 publicly available experimentally determined structures combined with an OpenFold-curated synthetic database containing more than 13 million structures. Training utilized up to 256 GPUs operating continuously for several weeks.
Sources
