Hugging Face releases Carbon-X, an open model that reads DNA

Hugging Face released Carbon-X on Thursday, an open 14.8 billion-parameter language model that takes raw DNA in an ordinary chat and answers in English, the company said on its model card. The weights are available under the Apache 2.0 license.

What it does: Hugging Face said Carbon-X predicts whether mutations are benign or pathogenic, identifies regions that control gene activity, classifies sequences by organism group, predicts the next bases of a sequence and explains its reasoning. DNA and text share a single token stream in one decoder, with no separate DNA encoder, and the model reads DNA in six-letter chunks, according to the model card.

Benchmarks: On five DNA benchmarks chosen by Hugging Face, Carbon-X averaged 66.7% with thinking disabled, 27 points above Qwen3-14B, the model it starts from, the company said. Hugging Face said that is “close to GPT-6.1 Sol (67.5%) at about 1/1,800 of the inference cost per problem” and that Carbon-X “leads every model we tested” on coding-variant effects and next-base prediction. The company estimated that analyzing the 2.4 million variants of uncertain significance in ClinVar would cost about $300 with Carbon-X, compared with $32,000 to $4 million with the frontier models it evaluated.

Built to run locally: The weights fit on a single 80 GB GPU, so researchers can keep sensitive or unpublished sequences on their own infrastructure, Hugging Face said. Lewis Tunstall, a senior research scientist at Hugging Face, told Wired that the team intentionally designed a small model so it could run on a user’s hardware, and that no data is sent to Hugging Face or any other company. “We’ve unlocked a new capability in language models, which is fundamentally about being able to chat with DNA,” Tunstall told Wired.

Training: Hugging Face said the model was trained on 302 billion tokens of DNA, biology text, English and instruction data. Tunstall told Wired the team did not train it on viruses because it did not want the model used for potentially dangerous applications.

Partners: Hugging Face said Carbon-X is a joint collaboration with research teams at Zhongguancun Academy in China and the Telethon Institute of Genetics and Medicine and the University of Naples Federico II in Italy.

Limits: Hugging Face said Carbon-X “has not been clinically validated” and should not be used to make medical decisions. It said the model can miscount bases, may invent a sequence’s identity or function without enough context, and remains weak on splice-site and histone-mark questions. The developers told Wired they have not yet validated the model with wet lab experiments, so they can’t say how accurate its predictions are.

Why it matters: Carbon-X pairs an open license with a model small enough for a single GPU, which addresses a constraint Tunstall described to Wired: researchers face restrictions on uploading sensitive patient data to frontier AI models. If Hugging Face’s cost estimates hold, large-scale screening of uncertain variants would be far cheaper than with the frontier systems it tested. Its chat interface could also open genomic questions to people without technical training, a use Tunstall pointed to. But the benchmark results are Hugging Face’s own, the model trails rivals on some tasks, and the company says it has not been clinically or experimentally validated, so its usefulness for real diagnoses or drug design remains unproven by the sources’ own account.

What’s next: Hugging Face said Carbon-X is a starting point for research on connecting DNA and language. Tunstall told Wired the tool is meant to help researchers design drugs and genetic therapies and could potentially aid patients in interpreting their own genomes.