Hugging Face has released Carbon-X, an open 14.8-billion-parameter model that lets researchers paste raw DNA into a chat and ask questions about it in plain English, such as whether a mutation is likely to cause disease.
The model, published on Hugging Face under an Apache 2.0 licence, was built with Zhongguancun Academy and the Telethon Institute of Genetics and Medicine at the University of Naples Federico II. Cableado reports it was released on Thursday, October 22.
Cómo funciona
Carbon-X starts from Alibaba’s Qwen3-14B. The team added DNA tokens that read sequences six bases at a time, then trained the model on 302 billion tokens of DNA, biology text, English text and instruction data, so DNA and language share one token stream.
You give it a reference and a mutated sequence, and it says whether the variant looks benign or pathogenic and explains why. It can also spot regions that control gene activity, classify which kind of organism a sequence comes from, and predict the next bases.
The numbers
According to the model card, Carbon-X averages 66.7% across five DNA benchmarks. That’s 27 points above the Qwen3-14B model it started from, ahead of every other open model Hugging Face tested, and close to GPT-6.1 Sol’s 67.5% at about 1/1,800 of the inference cost per problem.
Hugging Face estimates that checking the 2.4 million “variants of uncertain significance” in the ClinVar database would cost about $300 with Carbon-X, against $32,000 to $4 million with the frontier models it evaluated. The weights fit on a single 80GB GPU, so labs can keep sensitive sequences on their own hardware.
The limits
Hugging Face is clear that Carbon-X “has not been clinically validated, and it should not be used to make medical decisions.” It’s still weak on splice sites (29.8%) and histone marks (34.6%), can miscount bases, and may invent a sequence’s function when it lacks context. It wasn’t trained to design new DNA.
Lewis Tunstall, a senior research scientist at Hugging Face, told Wired the team left viruses out of the training data to avoid potentially dangerous uses. The developers also haven’t yet validated its predictions in wet-lab experiments.
Why it matters
Specialist tools like Pliegue alfa 3 have changed parts of biology, but they aren’t built for conversation. Carbon-X puts DNA analysis behind a chat window, and it’s small enough for a lab to run itself. Like Microsoft’s MAI-DxO this week, though, it’s a benchmark result until it’s proven against real biology.
