Hugging Face’s Carbon-X lets you ask questions about DNA in plain English

Hugging Face has released Carbon-X, an open model that reads raw DNA pasted into a chat and answers questions about it in English. Wired reported the launch; the model card lists the weights as Apache 2.0.

Hugging Face’s bio team built the model with partners at Zhongguancun Academy, a Chinese institution, and Italy’s University of Naples Federico II. According to its model card, it can handle questions on variant effects, regulatory elements, sequence classification and next-base prediction.

A 14.8B model that fits on one GPU

Carbon-X is built on Qwen3-14B. The card describes a hybrid tokenizer: ordinary text goes through byte-pair encoding, while DNA is cut into non-overlapping groups of six letters. The context window is 40,960 tokens. The authors say mid-training used 302B tokens, followed by supervised fine-tuning and two rounds of reinforcement learning with verifiable rewards.

The weights fit on a single 80 GB GPU in bfloat16, according to the card. That matters for the use case Hugging Face stresses. Lewis Tunstall, a senior research scientist at the company, told Wired the model was kept small so researchers can run it on their own hardware, which avoids sending sensitive patient data to a cloud service. The card adds that how genetic data is handled still depends on whoever hosts the model.

Loading requires Transformers v5 or later with trust_remote_code=True, and the model can be served with vLLM. The card says llama.cpp and other GGUF runtimes are not supported, even though a GGUF repository sits next to the main one. The release also includes examples and a ready-made skill for coding agents.

Benchmarks and a very low cost claim

On five DNA benchmarks with thinking turned off, Carbon-X averages 66.7%, the card says. That compares with 67.5% for the frontier model it lists, GPT-6.1 Sol, and the card says Carbon-X is ahead of the other open models tested. Its best results are on coding-variant effects and next-base prediction.

The authors also estimate an inference cost of about 1/1,800 of the frontier models per problem. By their estimate, working through the 2.4 million variants of uncertain significance in ClinVar would cost about $300. Both figures are the developers’ own and have not been independently checked.

What the developers say it cannot do yet

Wired noted that the team has not tested the model against wet lab experiments, so the accuracy of its predictions in a real biological setting is unknown. The model card goes further and says Carbon-X is not clinically validated and should not guide medical decisions.

The card also lists weak spots. Splice-site questions score 29.8% and histone-mark questions 34.6%. The model can miscount or misplace bases, and its quality drops on long DNA generation. Wired adds that users cannot feed in a whole genome: the model was trained on stretches of roughly 50,000 base pairs, so it looks at small regions at a time.

Training data covered mammals including humans, other vertebrates, invertebrates, plants, fungi and protozoa. Wired reports that viruses were left out on purpose, because Tunstall said the team wanted to reduce the chance of dangerous uses. Whether that exclusion holds up against determined misuse is not something either source tests.

Where it sits among DNA models

Carbon-X follows the earlier Carbon family from the same Hugging Face bio team, which came in 500M, 3B and 8B sizes under Apache 2.0 and was trained on genomic sequence alone to predict the next base. Carbon-X adds the conversational layer. Wired contrasts it with Google DeepMind’s AlphaGenome, which returns numeric predictions and needs more compute, and with Evo 2 from the Arc Institute and Nvidia, which can generate whole genomes.

The corporate backdrop is also new. Wired says Nvidia agreed in September to acquire Hugging Face for nearly $13 billion. Hugging Face has been releasing a steady stream of work in other fields too, such as a library of free simulation-ready 3D assets.

For researchers, the practical test is whether a 66.7% average on the team’s chosen benchmarks translates into useful calls on real variants, and whether the weak categories matter for their questions. Until outside groups run it on lab data, the numbers show what the model can do on benchmark tasks, not how reliable it is as a tool for drug or therapy design.