Hugging Face's Carbon-X Lets Researchers Chat With DNA

Hugging Face has released Carbon-X, an open 14.8-billion-parameter model that takes raw DNA inside an ordinary chat and answers in English. According to its model card, the model can predict “whether mutations are benign or pathogenic”, identify regions that control gene activity, classify sequences by organism group and predict the next bases of a sequence. It is free to use under the Apache 2.0 licence.

The pitch is a cheap, private alternative to frontier models for genomic research. Hugging Face says Carbon-X scores close to OpenAI’s GPT-6.1 Sol on its five DNA benchmarks at a tiny fraction of the cost, and that the weights fit on one 80 GB GPU, so labs can keep sensitive sequences on their own machines. All of those figures come from Hugging Face’s own testing, and the company says the model “has not been clinically validated” and should not be used for medical decisions.

Hugging Face says it nears GPT-6.1 Sol on DNA tasks

Across five benchmarks covering mutation effects, sequence understanding and next-base prediction, Carbon-X averages 66.7% with its reasoning mode switched off, Hugging Face says. That is 27 points above Qwen3-14B, the model it was built from, and just behind GPT-6.1 Sol’s 67.5% “at about 1/1,800 of the inference cost per problem”. It leads every model Hugging Face tested on two tasks, coding-variant effects and next-base prediction, while frontier models stay ahead on the other three.

Two of the five benchmarks, FrontierVEP and NextBase, were built by Hugging Face itself. The company also lists weak spots: it scores 29.8% on splice-site questions and 34.6% on histone marks, and it can miscount bases or invent a sequence’s identity when it lacks context. Hugging Face estimates that analysing 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. Its cost figure assumes an H200 GPU at $5 an hour.

The model was built with researchers at Zhongguancun Academy and the Telethon Institute of Genetics and Medicine at the University of Naples Federico II. It is a follow-up to Carbon-A, a 1.2B gene-finding model Hugging Face released earlier this month.

DNA and English share one model

Most genomic models read DNA but cannot explain themselves, while chatbots know biology mainly from text. Carbon-X puts both into a single decoder, the release blog says. Hugging Face added 4,096 six-letter DNA tokens to Qwen3’s vocabulary, then trained on 302 billion tokens, more than half of them raw DNA, followed by fine-tuning and reinforcement learning with rewards that check answers against the supplied sequence. The context window is 40,960 tokens.

“We’ve unlocked a new capability in language models, which is fundamentally about being able to chat with DNA,” Lewis Tunstall, a senior research scientist at Hugging Face, told WIRED. Hugging Face’s blog also claims that adding Carbon-X as a sub-agent raised Kimi K3’s score on its DNA index from 49.4% to 60.5%.

On misuse, the blog says the team excluded viral genomes that infect eukaryotes from its training data and found “limited editing reliability and no advantage over readily available deterministic tools” on sequence-editing tasks with dual-use relevance. It adds that more research is needed, including on what happens after fine-tuning.

The release documents disagree on where it runs

The most useful thing about Carbon-X may be its argument against relying on closed models for biology. Hugging Face’s blog shows frontier models refusing legitimate genomic requests more often as reasoning effort rises: Gemini 4 Argon’s FrontierVEP accuracy falls from 58% to 38%, and Claude Opus 5.5’s NextBase score drops from 17.7% to 1.1%. If those numbers hold up, a lab screening millions of variants has a practical reason to want an open model it controls, not just a cheaper one.

But the release materials tell conflicting stories, and that matters for anyone deciding whether to try it. The model card says Carbon-X needs a single 80 GB GPU and that llama.cpp and similar runtimes “are not supported” because the DNA tokeniser is custom code. The blog says it “works with llama.cpp” and shows it running on a MacBook. The blog’s opening line claims GPT-6.1 Sol-level results at “1,000× lower cost”, while the body and model card say 1/1,800. And the blog says it can “generate new DNA”, while the model card says it “was not trained to design new ones”. Researchers should treat the model card as the specification until Hugging Face reconciles them.

The benchmark story needs the same care. A model that is strongest on tests its own makers designed is a reason for others to check, not a settled result. Carbon-X joins a busy month for AI in biology, from OpenAI’s GPT-Rosalind agents to RFD4-Proteina and Radical Numerics’ claim of an AI-designed phage given to a patient. The test for Carbon-X is whether outside labs can reproduce its variant-effect scores on data Hugging Face did not choose.