Anthropic releases faster kernels for 36 biomolecular tools

Anthropic has released Apache 2.0 optimization code for 36 open-source biomolecular modeling tools used in protein structure prediction, drug design, and genomics. Across its benchmarks, Anthropic reports roughly 4x average acceleration in faster modes and about 1.6x acceleration with bit-identical outputs. A low-memory mode also lets researchers model systems above 10,000 tokens on a single NVIDIA GPU server.

Anthropic and Adaptyv Bio are also co-sponsoring a protein design competition detailed on the competition page. The program includes up to $1 million in Claude credits, $250,000 in Modal compute, DNA synthesis from Twist Bioscience, and wet-lab validation for more than 5,000 community designs.

Cubic geometry meets custom CUDA

Structure predictors such as AlphaFold3, OpenFold3, and Boltz-2 rely heavily on triangle attention and triangle multiplication. These operations update relationships among triplets of molecular tokens, which can represent amino acids, nucleotides, ligand atoms, or ions. Their naive implementations scale roughly as O(N³) in runtime and intermediate memory, so doubling the token count can increase both by a factor of eight.

Anthropic says Claude helped write custom GPU kernels for those bottlenecks. The resulting FlashPairformer implementation delivered 2.7x to 2.9x acceleration for triangle attention and 1.7x to 3.2x for triangle multiplication, depending on the model configuration. Anthropic reports that these kernels outperformed NVIDIA's cuEquivariance and BioNeMo Inference Runtime implementations in its tests.

Four modes, explicit failures

Model-specific changes supplement the shared kernels by caching repeated computation and replacing dead branches with their constant outputs. Anthropic says each optimized model was checked for downstream task quality. Every kit in the GitHub repository exposes four modes:

Every run prints one ACTIVE line identifying the enabled optimization. A mode that cannot engage on the current machine prints NOT ACTIVE and exits with status 3, preventing silent fallback to the stock implementation. Some kits also support --n_gpu P, which partitions one prediction across multiple GPUs in the same server.

Big mode crosses 10,000 tokens

Big mode lowers peak memory enough to model biomolecular systems above 10,000 tokens accurately, according to Anthropic. The infrastructure can also execute inputs above 70,000 tokens. A GPU node here means one server that may contain several local GPUs; Anthropic's largest disclosed test used eight B300 GPUs.

Successful Big-mode folds included human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial ribosome. Anthropic reports that each closely matched its experimentally determined structure. The 70S ribosome contains more than 10,000 tokens, while the 40S ribosome previously predicted accurately by AlphaFold3 contained 7,663.

The pipeline also processed viral capsids ranging from 31,000 to more than 70,000 tokens on an eight-GPU B300 node. Those runs completed but produced collapsed structures. Anthropic attributes the failures to a training-context ceiling, meaning the model had moved beyond the size range represented well enough during training.

The $150 computational benchmark

Anthropic also rebuilt the agentic binder-design pipeline from its earlier binder study on top of the optimized models. The revised setup sharply reduced prompt length, agent complexity, and compute:

Across 16 targets, the median-scoring and highest-scoring designs produced by three Claude models reached approximately the same ipSAE values as the earlier Mythos 5.1 campaigns while using about two orders of magnitude fewer GPU-hours. Anthropic reports that roughly $150 in combined GPU and token spending matched the earlier campaigns' computational performance. ipSAE estimates protein-interface quality and correlates with experimental binding; wet-lab testing remains a separate validation step.

Inside the 36-kit release

The repository ships one drop-in optimization kit for each pinned upstream tool, covering several modeling categories:

• Complex structure prediction: Boltz-2, Chai-1, Protenix, OpenFold3, RoseTTAFold3, and AlphaFold3 forks
• Binder design: BoltzGen, BindCraft, PXDesign, Genie 3, RFdiffusion 1, and RFdiffusion 3
• Inverse folding: ProteinMPNN and ESM-IF1, which generate sequences for target structures
• Protein language models: ESM-C, ProGen2, and E1
• Genomics: Evo 2, Enformer, Borzoi, ChromBPNet, and GPN-Star

Each kit sits beside a pinned copy of its upstream release and supports Docker, Apptainer, or a standard Python virtual environment. Existing upstream commands remain unchanged; an environment variable or --mode flag activates the optimization. The H100 80GB serves as the common reference configuration, with A100, H200, B200, and B300 configurations included where applicable.

Anthropic labels the repository as an unmaintained reference release. The pinned versions are provided as-is, with no pull-request intake or planned upstream updates. Apache 2.0 licensing permits adopters to fork and maintain the code independently.

Before putting it in a pipeline

Production evaluation needs to account for the difference between kernel benchmarks, complete workflow performance, and model quality on local data:

• Measure end-to-end wall time, including preprocessing, compilation, data transfer, and post-processing.
• Confirm that each run prints the expected ACTIVE mode on the target hardware and software stack.
• Revalidate task-specific metrics before adopting fast, even when numerical differences fall within reported seed variation.
• Validate Big-mode structures independently because successful completion establishes memory capacity rather than structural accuracy.
• Pin the kit commit and upstream model version, then plan for an internal fork if long-term maintenance is required.

Claude's role in the engineering

Anthropic says Claude completed the optimization work in under four weeks under the supervision of two staff members with biomolecular modeling expertise and no previous experience in inference optimization or GPU kernel engineering. The release provides inspectable kernels, model-specific patches, benchmark configurations, and downstream checks for evaluating that claim.

The four-week timeline defines a concrete use case for AI-assisted scientific software engineering: adapting shared low-level kernels across many specialized models under expert supervision. Reproducing the gains on additional hardware and newer upstream releases will show whether the same workflow can reduce the continuing cost of maintaining scientific computing infrastructure.