Hugging Face releases Carbon-X, an open model that lets you chat with DNA
Hugging Face on Thursday released Carbon-X, a new open model for understanding DNA that can run locally on a laptop. It was developed with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy. The model is trained on genomic sequences and natural language, so people can ask questions about the effects of DNA changes on gene activity and health and get conversational answers back. Wired describes it as something like a ChatGPT for DNA.
Lewis Tunstall, a senior research scientist at Hugging Face, says the tool is meant to help researchers design new drugs and genetic therapies, and that it could potentially help patients interpret their own genomes. He says the model unlocks a new capability in language models, which is fundamentally about being able to chat with DNA. Because it uses natural language, he says, even nontechnical users can ask questions about DNA that would not be accessible through any other tool.
The article sets Carbon-X against two earlier models. Google DeepMind's AlphaGenome, released last year, predicts the effects of DNA sequences and genetic mutations but gives its answers in a numeric format, and it needs more compute. Evo 2, developed by the Arc Institute and Nvidia along with academic partners, speaks the language of DNA and can generate entire genomes from scratch. Carbon-X, by contrast, combines DNA and natural language in a conversational interface.
On how it was built: the team started with an existing language model and expanded its vocabulary so it reads DNA in groups of six letters. They then continued training on 300 billion tokens that combined DNA sequences, English text, and examples linking DNA to biological descriptions. The DNA came from mammals (including humans), other vertebrates, invertebrates, plants, fungi, and protozoa. Tunstall says viruses were left out of training because the team did not want the model used for potentially dangerous applications.
There is a hard limit on input. The model was trained to ingest DNA sequences of roughly 50,000 base pairs, so users cannot upload a full human genome (about 6 billion letters) and can query only small parts of it at a time.
One suggested use is helping scientists work out whether a genetic mutation is benign or disease-causing, a key step in developing a targeted drug or genetic therapy. If an intervention targets the wrong mutation, it could be totally ineffective or, worse, cause unintended harm.
The small size was deliberate. Researchers face restrictions on uploading identifiable or sensitive patient data to frontier AI models, so Tunstall says his team designed a model with 14 billion parameters that runs on the user's own hardware instead of relying on cloud-based models. He says no data is sent to Hugging Face or any other company.
The developers caution that they have not yet validated Carbon-X with actual wet lab experiments, so they cannot say how accurate its predictions are.
The article also notes that Hugging Face, described as essentially GitHub for AI models, agreed in September to be acquired by Nvidia for nearly $13 billion.
Key facts
- Carbon-X is an open 14-billion-parameter model from Hugging Face that answers conversational questions about DNA and can run locally on a laptop.
- It was built with researchers at Zhongguancun Academy (China) and the University of Naples Federico II (Italy), by expanding an existing language model to read DNA in groups of six letters and training further on 300 billion tokens.
- Inputs are limited to sequences of roughly 50,000 base pairs, so users query small parts of a genome rather than uploading a full human genome.
- Because it runs locally, Tunstall says no data is sent to Hugging Face or any other company, which matters for sensitive patient data.
- The developers say the model has not been validated with wet lab experiments, so they cannot say how accurate its predictions are.
Why it matters
Most of the human genome is still not understood, and AI models are starting to help. The article contrasts Carbon-X with AlphaGenome, which gives numeric predictions and needs more compute, and with Evo 2, which speaks the language of DNA and can generate entire genomes. Carbon-X's pitch is different: plain-language questions and conversational answers, on a model small enough to run on a laptop. Tunstall frames it as a new capability in language models, being able to chat with DNA.
Who it affects
Tunstall says the tool is meant for researchers designing new drugs and genetic therapies, and could potentially help patients interpret their own genomes. Scientists weighing whether a mutation is benign or disease-causing are one named use case. Researchers who face limits on uploading identifiable or sensitive patient data to frontier AI models are the audience the small, local design was built for.
How to use it
You ask questions about DNA in natural language and get conversational answers. The model runs on your own hardware, and Tunstall says no data is sent to Hugging Face or any other company. Input is limited: it was trained on sequences of roughly 50,000 base pairs, so you can query only small parts of the genome at a time and cannot upload a full human genome. No license name, download link, or model repository details for Carbon-X are given.
How solid is it
The facts about the model's design come from Wired's report and from Tunstall, a senior research scientist at Hugging Face. The developers themselves say they have not validated Carbon-X with actual wet lab experiments and cannot say how accurate its predictions are. No benchmark results or accuracy figures are given. Claims about helping patients are stated as possibilities, not as demonstrated results.
Risks and caveats
Accuracy is unknown, so answers about mutations should not be treated as verified. The 50,000 base pair limit rules out whole-genome queries. On safety, the team left viruses out of the training data because they did not want the model used for potentially dangerous applications. The patient angle is only that the tool could potentially help people interpret their own genomes; the article does not say it is available to patients now.
“We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA”
— Lewis Tunstall, senior research scientist at Hugging Face