Model Book/Chai-1

Chai-1

B+

Chai Discovery

Structure PredictionComplex Predictionnon-commercial
Updated 1 month ago
NextIn take

Chai-1 proved that a small startup could match DeepMind on structure prediction accuracy — now the question is whether anyone will pay for a non-commercial license when Boltz-1 is fully open.

Specifications

ArchitectureDiffusion-based with enhanced multi-chain conditioning
Parameters~500M
Training DataPDB, UniRef, curated complex datasets
LicenseWeights available for non-commercial use. Commercial licensing requires agreement with Chai Discovery.
Hardware1x A100 40GB for most predictions
Inference CostFree via Chai Discovery platform (rate-limited), self-hosted for research
API AvailableYes
Weights AvailableYes

Benchmark Performance

BenchmarkScore
CASP150.83 GDT-TS (avg)

When to Use This

  • +Academic structure prediction research
  • +Benchmarking against AF3 and Boltz with accessible weights

When NOT to Use This

  • Commercial drug discovery without a licensing agreement
  • Production deployment where MIT-licensed alternatives exist

Production Readiness

research

Known Users

  • Chai Discovery internal
  • Academic research groups

Grade Rationale

B+

Strong accuracy that rivals Boltz-1 and approaches AF3, with available weights for research. The non-commercial license limits production adoption, keeping it below A-tier for teams that need to ship.

Sources

Update History

2026-03-10Initial entry