NVIDIA and Arc Institute unveil Evo 2 for genomics

Designing new biology has historically felt like an unpredictable guessing game where scientists spent months piecing together tiny DNA fragments.

Now, biomedical research organization Arc Institute and Stanford University have teamed up with NVIDIA to release Evo 2, the largest publicly available foundation model for genomic data.

The model is officially live today for researchers worldwide through the NVIDIA BioNeMo platform, offered as an easily deployable NIM microservice or as a full open-source download for custom fine-tuning.

Reading 9 trillion nucleotides with a million-token context

Evo 2 wasn't built like a typical biological model trained on narrow subsets. It was trained on nearly 9 trillion nucleotides spanning all domains of life, including plants, animals, and bacteria.

To make that possible, NVIDIA supplied an AWS cluster of 2,000 H100 GPUs via DGX Cloud, working side-by-side with researchers to scale and optimize the architecture.

The defining technical breakthrough here is sequence length. A single human gene can easily contain thousands of base pairs, and previous models struggled to see how distant mutations affected overall cell mechanics.

Evo 2 can process sequences up to 1 million tokens at once. That wide-angle lens allows it to link distant parts of an organism's genome directly to cellular function and gene expression.

In early benchmarks evaluating BRCA1, a gene notoriously linked to breast cancer, the model predicted whether previously uncataloged mutations would disrupt gene function with 90% accuracy.

Why it matters

We think Evo 2 marks a real shift from observing biology to generating it predictably.

For researchers in oncology, immunology, and agricultural tech, having a 1-million-token context window means you can feed entire genetic pathways into an AI instead of guessing how distant mutations interact. The availability of BioNeMo NIM microservices also means smaller biotech labs don't need millions of dollars in local supercomputers just to test a hypothesis.

For NVIDIA, this is another clear play to turn life sciences into its next major compute market. By bundling DGX Cloud, proprietary optimization, and BioNeMo deployment into one stack, they are locking in the biomedical research pipeline just like they did with generative text.

That said, high digital predictive accuracy is only the first hurdle. While a 90% hit rate on BRCA1 variants is impressive on a screen, validating those synthetic molecules in wet labs and clearing clinical trials will remain a slow, costly bottleneck for years to come.