Boltz
MIT-licensed co-folding models for structure and binding affinity prediction, plus BoltzGen for de novo binder design. Runs on a single GPU in roughly 20 seconds per prediction.
Overview
Boltz began as the open reproduction of AlphaFold 3-class co-folding out of MIT CSAIL and the Jameel Clinic. Boltz-1 landed in November 2024 under MIT, Boltz-2 followed with a preprint dated 14 June 2025 and added the thing structure prediction alone never gave you, binding affinity, and BoltzGen arrived in late 2025 (bioRxiv preprint 20 November 2025) with all-atom de novo binder design. Weights, inference code, training code and the designs themselves are all MIT, usable commercially without asking anyone.
The practical claim on Boltz-2 is FEP-class ranking at roughly 1000x less compute, turning a calculation that cost about $100 and 6 to 12 hours into one that costs cents and takes about 20 seconds on a single GPU. It exposes affinity two ways, a binary probability suited to hit discovery and a predicted value for optimization, which matters because those are different jobs with different error tolerances. Independent evaluation is mixed and honest people say so. On the four-target FEP+ subset it reaches about 0.66 Pearson against 0.78 for commercial FEP+ and 0.66 for OpenFE, while some external sets report much worse. Accuracy correlates with MSA depth, memory scales quadratically with sequence length and the 1,024 token region is where single-card runs start hurting.
Governance changed in January 2026 when the team incorporated as Boltz PBC, a public benefit corporation, with a $28M seed led by Amplify Partners alongside Andreessen Horowitz and Zetta Venture Partners, a Pfizer collaboration announced the same month and a later Takeda deal. Boltz Lab, the hosted platform, went to beta in January 2026 and the Boltz API followed on 16 June 2026 with BoltzMol-1 for small-molecule hit discovery and BoltzProt-1 for protein design. Those two are the first releases where hosted access came first, which is the tension to keep an eye on. For now the open lineage through Boltz-2 and BoltzGen is complete enough that a small team with one decent GPU can do real structural biology work without a vendor contract.
Key Features
- ✓ Boltz-2 predicts complex structure and binding affinity jointly, approaching FEP-class accuracy at roughly 1000x lower compute
- ✓ Two distinct affinity outputs: a binary probability for hit triage and a predicted value for lead optimization
- ✓ BoltzGen adds all-atom de novo binder design across peptides, proteins and other modalities, with weights, training code and every design released under MIT
- ✓ Runs locally on one CUDA GPU with a CPU fallback, installs with pip (package at 2.2.1), Python 3.10 to 3.12
- ✓ MSA server integration for input generation, plus batch processing for screening campaigns
- ✓ Optional hosted API since June 2026 for teams that would rather not own accelerators
Where it holds
- • Actually open, not open-ish. MIT covers code, weights and the training pipeline, so commercial use carries no licence conversation
- • Cost per prediction collapsed from around $100 and half a day of compute to cents and roughly 20 seconds on one GPU
- • Funded maintenance behind it: Boltz PBC launched January 2026 with a $28M seed from Amplify, Andreessen Horowitz and Zetta, plus Pfizer and Takeda collaborations
- • External benchmarking exists and is public, including Rowan Scientific's tracking of independent results, which is rare for models in this field
Where it breaks
- • Affinity numbers are noisy. On the canonical FEP+ subset Boltz-2 sits near 0.66 Pearson against 0.78 for commercial FEP+, and some external sets report poor or even negative correlation
- • Accuracy tracks MSA depth, so designed proteins without natural homologs come back low confidence. Membrane targets such as GPCRs and ion channels show high variance
- • Memory grows quadratically with sequence length and the practical ceiling sits near 1,024 tokens, which limits large complexes on one card
- • BoltzMol-1 and BoltzProt-1, the June 2026 pipelines, arrived as API products, so the open weight trail now lags the best hosted models
- • Small-molecule stereochemistry and allosteric binders are still weak spots the authors do not hide
My Take
Rare thing in this category: a model you can actually download, with MIT covering the weights and the training pipeline, from a team that has kept shipping since Boltz-1 in November 2024. Affinity prediction is both the headline and the caveat, since the FEP+ comparison reads as near-parity rather than replacement, roughly 0.66 Pearson against 0.78 on the canonical subset, and results wobble badly on membrane targets and allosteric sites. Use it as a fast triage filter ahead of physics, not as a substitute for it. The direction of travel is worth watching, because June 2026's BoltzMol-1 and BoltzProt-1 shipped as API products rather than checkpoints.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- MIT
- Starting at:
- Free. Code, weights, training pipeline, datasets and benchmarks are MIT licensed, so commercial use needs no agreement and no negotiation. Self-hosting costs whatever your GPU costs, with predictions in the region of cents rather than the ~$100 and 6 to 12 hours a comparable FEP calculation used to take. A hosted Boltz API launched 16 June 2026 from around $0.025 per prediction, and Boltz Lab, the managed platform in beta since January 2026, gives out monthly free credits.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Boltz holds and where it breaks. The fastest check is your own workload.
Visit website ↗