Hugging Face releases an open AI that reads DNA without data ever leaving the lab

Is this gene mutation harmless or pathogenic? That is the kind of question Carbon-X, Hugging Face's new AI model, is meant to handle by reading the DNA sequence directly and then explaining its reasoning in plain language.

The model, presented on October 22 by Wired, is openly available on the platform of the French-American startup under the Apache 2.0 license, which allows commercial use. It has about 14.8 billion parameters and, according to its model card, fits on a single 80 GB graphics card.

A language model that learned to read DNA

Carbon-X was not built from scratch. Its authors started from Qwen3-14B, a general-purpose model from China's Alibaba, and retrained it on about 302 billion text units: 55% raw DNA, 20% English text, 15% biology texts and 10% instruction examples. Supervised fine-tuning and a reinforcement learning phase followed.

The model card lists several uses: estimating whether a mutation is benign or pathogenic, spotting the regions that regulate gene activity, classifying a sequence by the group of organisms it comes from, and predicting the next bases. Hugging Face also presents it as a tool that an AI agent can call. According to the model card, Chinese and Italian research teams took part in the project: the Zhongguancun Academy, the Telethon Institute of Genetics and Medicine, and the University of Naples Federico II.

There are practical limits. The model was trained on chunks of up to about 49,000 base pairs and accepts no more than 40,960 units as input: it is therefore impossible to submit an entire human genome, and it must be queried region by region. Inputs much longer than its training examples have not been evaluated.

Scores measured by the authors themselves

On five DNA benchmarks, Carbon-X averages 66.7%, versus 67.5% for OpenAI's GPT-6.1 Sol, according to the model card. These figures were measured by the Hugging Face team, and two of the five tests were created by it. No independent evaluation is available yet.

The model card itself flags its weaknesses: low scores at spotting splice sites (29.8%), a key step in reading genes, errors in counting and coordinates, and made-up answers. Above all, it states that the model has not been clinically validated and must not be used for medical decisions. Wired points out that no laboratory validation has been done yet: that is what will show whether the tool truly serves research.

Why running it in-house matters in Europe

The most concrete argument for a European laboratory has less to do with performance than with how the model is deployed. Genetic data is among the special categories of data protected by the GDPR. An open model installed on an internal server makes it possible to analyze such data without sending it to an outside AI service, often an American one.

For teams that want to try it, the model card says the weights fit on a single 80 GB graphics card with the vLLM inference engine. The same precautions apply as with any language model: check its answers, especially when they concern precise positions in a sequence.

It remains to be seen what geneticists, who can test the model on their own data, will say, and what the first independent evaluations show.