Hugging Face-led team open-sources Carbon, a fast DNA foundation model with 393k bp context

Three checkpoints (500M, 3B, 8B) ship under Apache 2.0 with open weights, code, and data; a 6-mer tokenizer and a new loss drive the team’s cited ~275x throughput vs Evo2 while the 3B matches Evo2-7B’s win rate, trained on ~1T tokens, and they say a single GPU can process a human genome in under two days.

By · Published

Illustration of the Carbon DNA foundation model processing a DNA helix on a single GPU.

Three checkpoints (500M, 3B, 8B) ship under Apache 2.0 with open weights, code, and data; a 6-mer tokenizer and a new loss drive the team’s cited ~275x throughput vs Evo2 while the 3B matches Evo2-7B’s win rate, trained on ~1T tokens, and they say a single GPU can process a human genome in under two days.