Nvidia’s new DNA model learns what token prediction misses

Nvidia released JEPA-DNA, a genomic foundation model on Hugging Face for non-commercial research use.

The model combines masked token prediction with a latent-space prediction objective, building on DNABERT-2.

Nvidia says the hybrid approach could learn richer biological representations than token-only methods.

The launch reflects a broader shift toward architectures that may better handle biology and other structured data.