Model Book/Boltz-2

Boltz-2

A-

MIT

Structure PredictionMolecular DockingComplex Predictionopen
Updated 2 weeks ago
NextIn take

Boltz-2 unifies structure prediction and docking into one open model — the workflow simplification alone makes it worth evaluating, even before you factor in the accuracy improvements over Boltz-1.

Specifications

ArchitectureUnified diffusion model for structure prediction and molecular docking
Parameters~400M
Training DataPDB, PDBBind, UniRef, BFD — expanded ligand-protein complex data
LicenseMIT license. Fully open weights and training code.
Hardware1x A100 80GB recommended
Inference CostSelf-hosted — ~$0.03/prediction on cloud GPU
API AvailableNo
Weights AvailableYes

Benchmark Performance

BenchmarkScore
PoseBusters72% valid
CASP150.83 GDT-TS (avg)

When to Use This

  • +End-to-end structure prediction plus docking in a single pipeline
  • +Production workflows needing open-source ligand-protein modeling

When NOT to Use This

  • Pure protein structure prediction — Boltz-1 is lighter and nearly as good
  • When you need the absolute best docking accuracy regardless of openness

Production Readiness

pilot

Known Users

  • Early adopters in biotech
  • Academic drug discovery labs

Grade Rationale

A-

Extends Boltz-1 with integrated molecular docking in a single unified model. The first open model to seriously challenge AF3 across structure prediction and docking simultaneously. Still early but the trajectory is clear.

Sources

Update History

2026-03-25Initial entry