Hugging Face has released Carbon-X, an open-source AI model designed to answer natural-language questions about DNA. The model combines genomic sequences and English text in a conversational interface, so researchers can ask about how DNA changes may relate to gene activity and health.
Carbon-X is designed to run on a user’s own hardware. Its developers see potential uses in research on drugs and genetic therapies, while cautioning that the model’s predictions have not yet been checked through wet lab experiments.
Turning DNA questions into conversation
DNA carries information in four letters: As, Cs, Gs, and Ts. A human genome contains 6 billion of these letters, but scientists still do not understand how most of the genome works, even though it has been sequenced end to end.
Carbon-X aims to make some questions about DNA easier to ask. Rather than returning only a numeric prediction, it can respond conversationally to questions about sequences and biological descriptions. The team designed it to read DNA in groups of six letters, expanding an existing language model’s vocabulary for that purpose.
The model was trained further on 300 billion tokens combining DNA sequences, English text, and examples that connect DNA with biological descriptions. The DNA material came from mammals, including humans, as well as other vertebrates, invertebrates, plants, fungi, and protozoa. According to Lewis Tunstall, a senior research scientist at Hugging Face, the team excluded viruses because it did not want the model used for potentially dangerous applications.
A research aid with practical limits
One possible research task is assessing whether a genetic mutation may be benign or disease-causing. That distinction matters when researchers are developing a targeted drug or genetic therapy: an intervention aimed at the wrong mutation could fail or cause unintended harm.
Carbon-X is not built to process a complete human genome in one go. It was trained to ingest sequences of roughly 50,000 base pairs, so users can query smaller portions of the genome at a time. That makes it a tool for examining specific regions rather than a single conversational analysis of an entire genome.
The model’s interface may also make DNA questions accessible to people who do not use specialist tools. But a conversational answer should not be mistaken for a confirmed biological finding. The developers say they have not validated Carbon-X with actual wet lab experiments, so they cannot yet state how accurate its predictions are.
Local operation and sensitive data
Carbon-X has 14 billion parameters, a size its team chose so the model can run locally on a user’s hardware instead of depending on a cloud-based model. Tunstall says no data is sent to Hugging Face or another company when users run it this way.
That setup addresses a practical concern for researchers working with identifiable or sensitive patient data, who face restrictions on uploading such information to frontier AI models. Local operation can keep queries on the user’s hardware, while the model’s limited sequence capacity means researchers work with selected stretches of DNA.
How Carbon-X fits into DNA AI
Carbon-X joins other AI efforts aimed at understanding genetic sequences. Google DeepMind’s AlphaGenome predicts effects of DNA sequences and mutations in numeric form and requires more compute power. Evo 2, developed by the Arc Institute and Nvidia with academic partners, works with DNA and can generate entire genomes from scratch.
Carbon-X combines DNA and natural language instead, emphasizing a back-and-forth interface. That is a different way to interact with genomic analysis, but the value of its answers still depends on scientific validation. For now, its stated promise is to help users explore DNA-related questions locally, while experimental work is still needed to establish how reliably its predictions reflect biology.