Anthropic open-sources 36 inference optimization kits for protein and genomics models, ~4x average speedup in under four weeks
Anthropic published 'How Claude is uplifting biomolecular modeling' on 2026-09-17 alongside the anthropics/uplifting-biomolecular-modeling repo, which holds 36 drop-in optimization kits for open-source structure prediction, protein design, protein language and genomics models including AlphaFold 3 (JAX and Torch), Boltz-2, OpenFold3, ColabFold, ESM C and RFdiffusion3. Claude, working inside Claude Science, optimized more than 30 models in under four weeks for roughly 4x average speedup, plus a low-memory mode that predicts systems over 10,000 tokens on a single NVIDIA GPU node. Each kit pins a stock upstream version next to an opt package and offers off/exact/fast/big modes, with a shared runtime carrying FlashPairformer and triangle-attention kernels; Apache 2.0, explicitly unmaintained. Anthropic is also co-sponsoring an Adaptyv Bio protein design competition backed by up to $1M in Claude credits and wet lab validation for over 5,000 designs.
Source
↳ Follow the thread