Claude speeds up 30+ open-source biology models
Anthropic says Claude optimized more than 30 open-source biomolecular models in under four weeks, averaging 4x speedups and cutting the GPU cost of its earlier protein design work roughly 100-fold.
Anthropic says an internal, general-purpose research model optimized more than 30 open-source deep learning models for biology — covering structure prediction, protein design, genomics, and protein language models — in just under four weeks, averaging a 4x speedup with minimal precision loss and nearly 2x with identical outputs. The company says it verified that the accelerated versions did not degrade downstream task performance, and is open-sourcing the optimized code alongside a technical report.
The core of the work is FlashPairformer, a set of custom kernels for triangle attention and triangle multiplication, the cubic-cost operations that dominate runtime and memory in models such as AlphaFold3, OpenFold3, and Boltz-2. Anthropic claims it beats the field standard — NVIDIA's cuEquivariance and BioNeMo Inference Runtime — by 2.7–2.9x on triangle attention and 1.7–3.2x on triangle multiplication, depending on configuration. A new low-memory mode allows prediction of systems larger than 10,000 tokens on a single NVIDIA GPU node.
Applied to Anthropic's earlier de novo binder design work — which consumed up to $10,000 per target on Modal, roughly 2,500 H100 GPU hours — the optimizations reportedly reach comparable in silico performance using two orders of magnitude fewer GPU hours.
Anthropic is also co-sponsoring a protein design competition with Adaptyv Bio across five selected problems, committing up to $1 million in Claude credits and $250,000 in Modal compute credits, with wet lab validation for over 5,000 designs; Twist Bioscience also contributed.
The results are first-party and, apart from the competition, in silico.