Hugging Face released a new open-source model on Thursday that is built to interpret DNA and respond to questions about it in ordinary language, according to WIRED. The model, called Carbon-X, was developed with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy, and it is small enough to run locally on a laptop.
Hugging Face is widely known as a repository where developers and researchers find and run ready-made AI tools, and the company says the new model is trained on both genomic sequences and natural language. Users can ask how a particular DNA change affects gene activity or health, and the model answers in conversation rather than in raw numbers.
Why it matters
Much of the human genome remains poorly understood even though it has been sequenced in full, WIRED reports. Carbon-X is aimed at narrowing that gap by making DNA queries accessible to people without specialist training.
Lewis Tunstall, a senior research scientist at Hugging Face, told WIRED the tool is intended to support the design of new drugs and genetic therapies, and that it may eventually help patients make sense of their own genomes. He described the release as a new capability for language models: the ability to hold a conversation about DNA. Because the interface uses everyday language, he said, users who are not technical can ask questions about DNA that other tools would not let them reach.
The model also arrives at a moment when AI is becoming central to genomics. Google DeepMind released AlphaGenome last year, which forecasts the impact of DNA sequences and mutations but returns results in numeric form and needs more computing power. Evo 2, built by the Arc Institute and Nvidia with academic partners, works in the language of DNA and can produce whole genomes from scratch. Carbon-X sits between those approaches, pairing DNA with natural language in a chat-style interface — something like a ChatGPT for genetic code.
How the model was built and what it will not do
DNA carries information in a four-letter alphabet — A, C, G and T — and the human genome holds roughly 6 billion of those letters, WIRED notes. The Carbon-X team began with an existing language model and widened its vocabulary so it reads DNA in blocks of six letters. Training then continued on 300 billion tokens combining DNA sequences, English text and examples that tie DNA to biological descriptions. The sequences were drawn from mammals including humans, other vertebrates, invertebrates, plants, fungi and protozoa.
Tunstall told WIRED the team deliberately left viruses out of the training data to avoid the model being turned to dangerous uses. There are other limits. Users cannot upload a complete human genome; the model was trained to take in sequences of about 50,000 base pairs, so queries cover only small stretches of the genome at a time.
The size of the model was also a deliberate choice. Researchers are restricted in what identifiable or sensitive patient data they can send to frontier AI systems hosted in the cloud, so the team built a compact model with 14 billion parameters. That lets it run on the user's own hardware, and Tunstall said no data is transmitted to Hugging Face or any other company.
One anticipated use is judging whether a genetic mutation is harmless or tied to disease — a step that matters when developing a targeted drug or gene therapy, since an intervention aimed at the wrong mutation may do nothing or cause unintended harm.
The developers have acknowledged a significant caveat: Carbon-X has not yet been checked against actual laboratory experiments, so they cannot say how reliable its predictions are. WIRED also noted that Hugging Face agreed in September to be acquired by Nvidia for close to $13 billion, and that the company was at the center of a major hack involving OpenAI's models earlier this year.
What to watch: whether independent wet-lab studies validate the model's predictions, and how researchers use it in drug and gene therapy work.
