BoltzMol-1 hit on 6 out of 10 previously unseen targets while screening only 28-51 compounds per target. That’s a genuinely small number. Traditional screening campaigns can run thousands of compounds per target just to find a handful of hits. They’re compressing that dramatically.

BoltzProt-1 nearly tripled hit rates on protein design compared to their earlier version, with 58% of candidates passing developability filters. That second number matters a lot in practice because getting a hit that immediately fails ADMET or expression is not actually useful.

They also launched a GPU API at $0.025 per prediction built for agentic workflows, which suggests they’re thinking about how this plugs into automated pipelines rather than just being a one-off tool.

Excited to see how this holds up across more diverse target classes. The sample sizes here are still small, but the direction is real.

Boltz is building in public and moving fast. This is what the next generation of drug discovery infrastructure looks like.

Boltz