Hugging Face releases Carbon-X, a 14.8B open model for chatting with DNA

Hugging Face's Carbon-X is the most capable model in its Carbon biology family. It lets researchers ask natural-language questions about genetic sequences and generate DNA. It was built with Zhongguancun Academy and the University of Naples Federico II.

The model has 14.8 billion parameters, is licensed under Apache 2.0, and fits on a single 80GB GPU. Coverage also describes local laptop use. On DNA benchmarks it scores 66.7%, close to GPT-6.1 Sol's 67.5%, at an estimated 1,800x lower inference cost. It can also plug into agent frameworks for automated sequence interpretation.

Local execution matters in genomics. Researchers handling patient sequences often cannot send data to cloud APIs, and open weights make the model auditable. The release continues the strong-small-open-model pattern that IBM's CEO cited when saying enterprises are choosing open weights over frontier APIs.

The main caveats are that benchmark parity does not guarantee clinical-grade reliability, and that generative DNA models raise biosecurity questions. Independent evaluation and responsible-use guidance will be worth watching.