Anthropic used Claude to optimize over 30 open-source biomolecular models. It achieved a 4x average speed-up with minimal precision loss. A low-memory mode now supports systems over 10,000 tokens on one GPU. It open-sourced all optimized code. Anthropic and Adaptyv Bio are co-sponsoring a protein design competition. The competition offers up to $1 million in Claude credits. It also includes wet lab validation for over 5,000 designs. Claude developed FlashPairformer kernels for faster triangle attention and multiplication. These kernels outperform the field standard by 2.7-2.9x on attention. They also outperform by 1.7-3.2x on multiplication. The optimizations were completed in under four weeks. Claude now designs de novo protein binders with two orders of magnitude fewer GPU hours. Performance matches earlier campaigns using far more resources.
What changed
Biomolecular models ran slower and required more memory, limiting system size and accessibility.
What it unlocks
Running large biomolecular models on a single GPU and designing proteins with far fewer GPU hours.
- 4x average speed-up for 30+ biomolecular models
- 2x speed-up with identical outputs
- 10,000+ tokens on a single NVIDIA GPU node
- $1M in Claude credits for competition
What you need to act on it
- access to optimized code
- a single NVIDIA GPU node for large systems
Sources