Hugging Face Launches Carbon-X, an Open-Source DNA Model Based on Qwen3-14B

Hugging Face

Carbon-X integrates DNA sequences and natural language into a single model: researchers can ask questions about specific DNA sequences, predict the effects of mutations, and generate new DNA sequences. It can also act as a specialized sub-agent, helping other models interpret DNA sequences that they themselves find difficult to understand.

Built on the Qwen large model

Hugging Face senior research scientist Lewis Tunstall said that Carbon-X is a first of its kind, combining DNA analysis and generation capabilities with natural language, enabling researchers to ask questions about DNA sequences and essentially "talk to DNA."

Hugging Face developed Carbon-X in collaboration with research laboratories in Italy and China, and it is an open-source model. Hugging Face developed Carbon-X by training the existing language model Qwen3-14B on genomic sequences and natural language. The model was trained on DNA sequences from humans, other mammals, plants, and fungi.

Hugging Face said that on five DNA understanding and generation benchmarks, Carbon-X outperformed all other open-source models the company evaluated, with an average score close to GPT-6.1 Sol, but at only 1/1800 of the latter's inference cost per question. Compared with Qwen3-14B, Carbon-X's score improved by 27 points, indicating that the performance of open language models can be improved by training directly on DNA data.

Carbon-X's benchmark scores approach those of GPT-6.1 Sol

Speaking about Carbon-X's application scenarios, Tunstall gave an example: researchers can provide the model with a DNA sequence and ask whether a particular mutation would increase a person's risk of disease. The model then gives a prediction and provides an explanation. It can also generate new DNA sequences, thereby helping researchers develop new gene therapies.

Preventing over-reliance on big companies

Tunstall said that part of Hugging Face's motivation for developing Carbon-X was to prevent biology researchers from becoming dependent on large AI laboratories. He noted that AI has already had a disruptive impact on the field of mathematics.

"If you want to truly push the frontier of mathematical research, you very likely have to work at an AI lab like OpenAI or Anthropic, because they have the best mathematical models," Tunstall said. He believes similar situations could arise in other scientific fields.

He added: "We want to provide tools to the open research community so that they do not have to rely entirely on a single external vendor."

Researchers can run Carbon-X on their own computers without sharing sensitive patient data with external AI companies. Tunstall also said that some of the misaligned behaviors observed in large models have not yet appeared in smaller, more specialized models, such as the problems exposed by the well-known incident in July this year when an OpenAI agent breached Hugging Face.

Tunstall said that general-purpose chatbots sometimes also refuse to answer legitimate biology questions posed by researchers because they mistakenly believe the questions are attempts to "design biological weapons." He said that to prevent misuse of the model, Carbon-X's training data explicitly excludes virus-related data.

Although Carbon-X is likely to be used mainly by researchers, Tunstall said the model is a small step toward the future, when AI is expected to help provide more personalized medical services.

"I think the ultimate goal is to achieve personalized medicine, where each of us can use our own medical records and genomic data to ultimately find and perhaps even design treatment plans that suit us," he said.