Hugging Face Releases Carbon-X, an Open-Source AI Model That Lets You Chat With DNA
Hugging Face released Carbon-X on October 22, an open-source AI model that lets researchers ask questions about DNA sequences in natural language and receive conversational answers. Users paste a genetic sequence, pose a question, and the model explains potential effects on gene activity and health. Lewis Tunstall, a senior research scientist at Hugging Face, described the capability as "chatting with DNA."
Carbon-X starts from Alibaba's Qwen3-14B language model and has roughly 14.8 billion parameters. Hugging Face extended its vocabulary so the model reads DNA in six-base groups (6-mers), then continued training on 302 billion tokens mixing DNA sequences, English text, and biology material. This lets the model treat genetic code and ordinary language as one continuous stream rather than two bolted-together systems.
Tunstall said the team intentionally built a small model so it can run on the user's own hardware. A single 80 GB GPU is enough, and no data is sent to Hugging Face or any other company. The design addresses restrictions researchers face when uploading identifiable or sensitive patient data to frontier AI models hosted in the cloud.
On Hugging Face's own tests across five DNA benchmarks, Carbon-X averaged 66.7%, slightly behind OpenAI's GPT-6.1 Sol at 67.5%. Hugging Face said it reaches that result at roughly 1/1,800 of the inference cost per task. These figures are company claims that have not been independently verified, and the benchmark names are not specified in publicly available materials.
Carbon-X was developed in collaboration with researchers at Zhongguancun Academy in China and the University of Naples Federico II and the Telethon Institute of Genetics and Medicine in Italy. Training data spans DNA sequences from mammals (including humans), other vertebrates, invertebrates, plants, fungi, and protists. Viruses were explicitly excluded.
Tunstall noted that generalist chatbots sometimes refuse legitimate biology questions because they mistake them for attempts to design bioweapons. To avoid misuse, the team left viruses out of Carbon-X's training data entirely. The model ships under the Apache 2.0 license and includes a ready-made skill for coding agents such as Claude Code and Codex CLI.
Scientists could use Carbon-X to help determine whether a genetic mutation is benign or causes disease, a key step in developing targeted drugs or gene therapies. The model can also explain variant effects, regulatory regions, and organism groups. However, Hugging Face stated Carbon-X is not clinically validated and has not been tested in wet-lab experiments, so it should not be treated as a medical diagnostic tool.