Loading page…
Carbon-X combines genetic sequences with conversational answers and keeps local users’ data on their hardware, its developers say. Its prediction accuracy remains unconfirmed by laboratory experiments.
Listen to this story
Carbon-X’s key distinction is that it turns DNA analysis into a conversational interface rather than returning only numerical predictions. Hugging Face says the 14-billion-parameter model can run locally, addressing concerns about sending sensitive genetic sequences to cloud services, but each query is limited to about 50,000 base pairs. The developers have not tested its predictions in laboratory experiments, so its accuracy—and its value for interpreting mutations or guiding treatment—remains unestablished.
Training used 300 billion tokens combining DNA, English, and paired genetic and biological descriptions; the model reads DNA in six-letter groups.
The team excluded viruses from training, with senior research scientist Lewis Tunstall citing a desire to avoid potentially dangerous applications.
Carbon-X differs from AlphaGenome’s numerical predictions and Evo 2’s genome generation by pairing DNA sequences with natural-language questions.
Hugging Face released Carbon-X on October 22, 2026, an open-source model that lets people ask questions about DNA in ordinary language. The system is designed to run locally, bringing conversational access to genetic analysis. But its developers say they have not validated its predictions in laboratory experiments and cannot yet say how accurate they are.
Developed with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy, Carbon-X combines genomic sequences with natural-language text. Users can ask about how changes in DNA might affect gene activity and health, receiving conversational answers rather than only numerical predictions.
The team began with an existing language model and extended its vocabulary to read DNA in groups of six letters. It then continued training on 300 billion tokens—chunks of data—mixing DNA sequences, English text and examples that connect genetic sequences to biological descriptions. That pairing underpins the model’s ability to discuss DNA in words.
Its training sequences span mammals, including humans, as well as other vertebrates, invertebrates, plants, fungi and protozoa. The team deliberately left out viruses, Hugging Face senior research scientist Lewis Tunstall told WIRED. He said that choice was intended to avoid potentially dangerous applications.
The conversational interface distinguishes Carbon-X from other DNA models described in WIRED’s coverage. Google DeepMind’s AlphaGenome predicts the effects of sequences and mutations in numerical form. Evo 2, developed by the Arc Institute, Nvidia and academic partners, can generate whole genomes. Carbon-X instead brings DNA and natural language together in a chat interface.
Carbon-X was trained to ingest sequences of roughly this length. Users cannot upload a full human genome.
Carbon-X has 14 billion parameters, the internal values a model learns during training. Tunstall said the team intentionally kept it small enough to run on users’ own hardware. That design responds to restrictions researchers face when uploading identifiable or sensitive patient information to frontier AI models.
With local operation, no user data is sent to Hugging Face or another company, Tunstall said. The model therefore does not require researchers to send those sequences to a cloud-based service to obtain an answer. Its input limit still confines each query to a small portion of a genome, rather than a person’s complete genetic sequence.
One proposed research use is assessing whether a genetic mutation is harmless or disease-causing. That distinction matters when selecting a target for a drug or genetic therapy. An intervention aimed at the wrong mutation could fail to work or cause unintended harm, making the biological interpretation consequential beyond the chat interface.
Tunstall describes Carbon-X as a tool intended to help researchers design drugs and genetic therapies. He also sees a potential role for patients trying to interpret their own genomes. Those ambitions extend its audience beyond specialists: he argues that natural-language questions can make DNA analysis accessible to people without technical expertise. Patient interpretation remains a proposed use, not a demonstrated health benefit.
Loading discussion...
Join the conversation
Explain where broader access would help, or where it would create risk.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.