Hugging Face's Carbon-X lets you ask a DNA sequence questions in plain English, and it runs on a laptop

Hugging Face on Thursday, Oct. 22, 2026, released Carbon-X, an open-source model that reads both DNA and natural language, so researchers can ask what a sequence does or whether a mutation is likely harmful and get a conversational answer, running locally on their own hardware.

Most DNA models talk in numbers; most chatbots talk about DNA without really reading it. Carbon-X tries to do both, and the pitch that matters is that it runs locally, which is the only way a lot of hospitals and labs are ever allowed to put real patient sequences into an AI. The benchmark claims are Hugging Face grading its own homework, and nobody has checked the predictions at a lab bench yet, so “ChatGPT for your genome” is still a long way off. But an open, free model that gets near the frontier on mutation calls is exactly the kind of thing that ends up quietly inside a lot of research pipelines.

Carbon-X is an open model for understanding DNA and for generating it, and it also reads ordinary language. It has 14 billion parameters. A parameter is one number the model learned in training. The license is Apache 2.0, which lets anyone use, change, and share the model, including in a product they sell. A researcher can ask what a stretch of DNA does, ask how a mutation might change that, or ask the model to propose a new sequence. Another model can also call Carbon-X as a specialist and take the answer back. Those lines are Hugging Face's.

Hugging Face started from Alibaba's Qwen3-14B, an open model of the same size, and trained further on 300 billion tokens, a mix of DNA and ordinary text. A token, here, is one chunk of that training material. The DNA was read six letters at a time. After that came supervised fine-tuning, a pass on examples that already have answers, and reinforcement learning, a pass that rewards better answers. The work was done with researchers at Zhongguancun Academy in China and at the University of Naples Federico II and TIGEM in Italy. Those lines are Hugging Face's.

The scorecard is Hugging Face's own, across five DNA benchmarks. Carbon-X beat every other open model Hugging Face tested. The average was 66.7, against 67.5 for OpenAI's GPT-6.1 Sol. Hugging Face says running one of those problems costs about 1,800 times less. The same card puts Carbon-X 27 points above the Qwen3-14B model it started from. On coding-variant effects, a DNA change that alters the protein a gene codes for, and on next-base prediction, guessing the next letter, Hugging Face says Carbon-X leads every model it tested, open or closed. Frontier models still win on some tasks, including complex variant effects and the effect of changing a single letter. Those comparisons are Hugging Face grading its own tests.

The model is built to run on the user's own machine, so a sensitive sequence does not have to leave the building. Hugging Face's demo shows it on a laptop through llama.cpp, software that runs this kind of model locally. Lewis Tunstall of Hugging Face told WIRED that no data is sent to Hugging Face or to anyone else. A question can include roughly 50,000 base pairs at a time. A base pair is one letter of DNA. That cap is why a user is not uploading a whole genome. Those lines are Hugging Face's and WIRED's.

WIRED reports that the team deliberately left viruses out of the training data, to limit dangerous uses. Tunstall told WIRED, “We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA.” He said it could help researchers design drugs and gene therapies, and that it could potentially help patients interpret their own genomes. Those are his claims. WIRED and The Next Web both add the limit: the predictions have not been checked in a wet lab, meaning experiments with real biological samples rather than a computer, and Carbon-X is not a clinical tool.

WIRED also notes, briefly, that Nvidia agreed in September to acquire Hugging Face for nearly $13 billion. That is context for who would own the lab. It is not a term of this model release.

The picture is a frame from Hugging Face's official demo: a terminal running Carbon-X locally, answering “What does this sequence encode?”

In plain terms, Hugging Face on Thursday released a free, open model that reads DNA and plain English together and runs on a person's own computer. On Hugging Face's own tests it nearly matches a much costlier OpenAI model and beats the other open models they tried. Nobody has checked those predictions at a lab bench, and it is not a clinical tool. The training data left viruses out.