Anthropic open-sources 36 speedup kits for protein models

Anthropic open-sourced 36 inference optimization kits for protein and genomics models on GitHub, covering AlphaFold 3, Boltz-2, OpenFold3, ColabFold, ESM C and RFdiffusion3. For anyone running these six models, that means faster inference. The kits promise about 4x average speedup and let predictions of systems over 10,000 tokens run on a single GPU node, Anthropic says.

Claude did the optimization work inside Claude Science, tuning more than 30 models in under four weeks. Each kit pairs a stock upstream version of a model with a separate optimization package, so the original code stays untouched. The optimization package runs in four modes: off, exact, fast and big, and every kit shares a runtime built around FlashPairformer and triangle-attention kernels.

The repo is Apache 2.0 licensed and unmaintained, so bug fixes and compatibility work for future model versions fall to whoever adopts a kit.

Separately, Anthropic is co-sponsoring an Adaptyv Bio protein design competition. It's backing the contest with up to $1M in credits and wet lab validation for more than 5,000 designs.