Hugging Face Bio has introduced “Carbon,” a new family of open-source generative DNA foundation models designed to analyze and interpret genetic sequences. Trained on approximately 1 trillion tokens derived from eukaryotic DNA and RNA, the models can generate sequences, predict variant impacts, and identify functional motifs. The lineup includes Carbon-3B and Carbon-A, with the latter powering the Carbon Annotation Database containing over 566 million predicted protein-coding loci. Released under a permissive Apache-2.0 license on the Hugging Face Hub, the models offer high throughput and long context windows to lower computational barriers for researchers. Independent wet-lab validation has already backed hundreds of previously unannotated gene candidates identified by the models across various species.

Full report : Hugging Face releases Carbon-X, an open-source, 14 billion-parameter multimodal AI model to let users query DNA sequences via natural language, trained on 300 billion tokens.