Hugging Face released an open model on Thursday that lets users ask questions about DNA and receive conversational answers, marking what the company calls a new ability to "chat with DNA." Built with researchers from Zhongguancun Academy in China and the University of Naples Federico II in Italy, the model—called Carbon-X—is trained on both genomic sequences and everyday language, and it's designed to help researchers create new drugs and genetic therapies. Unlike other DNA models, this one can run locally on a laptop, keeping patient data private.
The team behind Carbon-X began with an existing language model and taught it to read DNA in chunks of six letters at a time. They then trained it on 300 billion tokens that blended DNA sequences, English text, and examples connecting DNA to biological explanations. The genetic material came from mammals—including humans—plus other vertebrates, invertebrates, plants, fungi, and protozoa. The model was intentionally not trained on viruses to prevent misuse for dangerous purposes. Users can query sequences of roughly 50,000 base pairs, which means they'll analyze small portions of a genome rather than uploading a full human genome at once.
According to Lewis Tunstall, a senior research scientist at Hugging Face, "We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA." He says the tool could help even nontechnical users "ask questions about DNA that wouldn't be accessible through any other tool," potentially allowing patients to interpret their own genomes. The developers intentionally built a small model with 14 billion parameters so it can run on local hardware rather than cloud-based systems, meaning no data is transmitted to Hugging Face or any other company. The team warns, however, that they haven't validated Carbon-X with actual laboratory experiments yet, so the accuracy of its predictions remains unknown.
One key application involves determining whether a genetic mutation is harmless or causes disease—a critical step when developing a targeted drug or gene therapy. If a treatment targets the wrong mutation, the report notes, it could fail completely or cause unintended harm. Carbon-X differs from earlier models like Google DeepMind's AlphaGenome, which predicts the effects of DNA sequences and mutations but outputs results in numeric format and requires more computing power, and the Arc Institute and Nvidia's Evo 2, which speaks the language of DNA and can generate entire genomes from scratch but doesn't offer a conversational interface. By combining DNA understanding with natural language, Carbon-X aims to make genetic analysis accessible to researchers who face restrictions on uploading identifiable or sensitive patient information to frontier AI models.
The model represents an attempt to democratize genomic research at a moment when scientists still don't understand how most of the human genome—all 6 billion letters of it—actually functions. The conversational design and local deployment could expand access beyond specialists, though the lack of wet lab validation means researchers will need to verify Carbon-X's suggestions before acting on them. For companies and labs navigating privacy regulations around patient data, the ability to query DNA locally without sending information to external servers may prove as valuable as the model's analytical capabilities themselves.