NederlandsHugging Face releases open AI model that lets researchers 'talk' to DNA
With Hugging Face's Carbon-X, researchers paste a piece of DNA into a chat and ask in plain language whether a change is likely to be harmful. The model is free and runs on a researcher's own computer. It is meant for research; the next step is testing in the lab.
On 22 October 2026 Hugging Face released Carbon-X, an AI model that lets researchers 'chat' with DNA. You paste a piece of genetic material into the conversation and ask a question about it in plain language. The model answers and explains how it reached that answer. Anyone may download and use it for free, also for commercial purposes.
Why reading DNA is so difficult
DNA is a long text in an alphabet of four letters: A, C, G and T. Every person has millions of small differences in it compared with other people, a kind of typing error. Almost always they are harmless. Sometimes such a change increases the risk of a disease. A single change can even cause a disease, as with Wilson's disease. Working out which change matters is a lot of work for researchers and doctors. The same goes for the question of which pieces of DNA decide whether a gene is switched on or off.
With Carbon-X, researchers can ask such questions directly. What does this piece of DNA probably do? Is this change probably harmless or harmful? Which parts control the activity of a gene? According to Lewis Tunstall, a researcher at Hugging Face, the model is meant to help research into new medicines and gene therapy. Gene therapy is a treatment that repairs an error in the DNA. In the long term, Tunstall sees the model as a small step towards personalised care. People could then use their own medical and genetic data to look for a treatment that suits them.
Built on an existing language model
Hugging Face is best known as the platform where researchers and companies share AI models with each other, but it also makes its own models. In September Nvidia announced that it wants to take over the company. For Carbon-X, Hugging Face worked with the Telethon Institute of Genetics and Medicine and the University of Naples Federico II in Italy, and with the Zhongguancun Academy in China.
The team started with an existing language model from the Chinese company Alibaba. That model learned a kind of extra vocabulary: it reads DNA in 'words' of six letters. It was then trained further on about 300 billion small pieces of DNA and text. Just over half of that was DNA, from humans, other animals, plants and fungi. The rest consisted of English texts about biology and example conversations. DNA from viruses was deliberately left out, so that the model is harder to use for harmful purposes.
Runs in the researchers' own lab
Carbon-X is small enough to run on a single powerful graphics card with 80 gigabytes of memory, of the kind research institutes often have. That is a deliberate choice. Patient data are usually not allowed to go to the servers of an AI company. With Carbon-X they stay in the lab. Tunstall points to another advantage: general chatbots sometimes refuse ordinary questions about biology, because they look like requests that could lead to misuse. A model made especially for researchers helps them move forward.
The model looks at pieces of DNA of about 50,000 letters at a time. The complete DNA of a human has more than three billion, so researchers always ask their questions about a small part of it.
Almost as good as the largest models
Hugging Face tested Carbon-X on five sets of questions about DNA. On average the model answered 66.7 per cent correctly, just below GPT-6.1 Sol from OpenAI (67.5 per cent) and above Claude Opus 5.5 from Anthropic (60.9 per cent) and Gemini 4 Argon from Google (49.8 per cent). The language model the team started with scored about 40 per cent. According to Hugging Face, a question to Carbon-X costs about 1,800 times less than the same question to GPT-6.1 Sol.
These figures come from Hugging Face itself, and the company also made two of the five test sets. Independent researchers can now repeat the results, because the model is freely available. On some types of question Carbon-X still scores around 30 per cent; there is room for improvement in future versions.
Next step: testing in the lab
Carbon-X is a research tool and has not been tested for medical use. According to Hugging Face it is therefore not suitable for decisions about patients. The model can also make mistakes. For example, it can count letters wrongly, or make something up about where a piece of DNA comes from if it gets too little information. The predictions have not yet been checked with experiments in the lab. That is the next step to see how reliable they are.
What does this mean for the road to the singularity?
Carbon-X shows that a language model can learn not only human language but also the 'language' of life, and that it can combine both in one conversation. That makes genetic knowledge more accessible. A researcher who is not a geneticist, for example a paediatrician with a patient who has a rare condition, will be able to ask in plain words what a particular DNA change might do. The model links that question to what it has learned from an enormous amount of biology texts.
Because the model is free and runs on a researcher's own computer, smaller labs and hospitals can use it too. According to Hugging Face, checking a large public database of 2.4 million known DNA changes with Carbon-X costs a few hundred dollars, instead of tens of thousands with the most expensive models. If independent tests and lab experiments show that the predictions are right, this kind of AI can make genetic research faster and cheaper, and so shorten the road to better treatments.
What is the technological singularity? Read the explanation →
