Hugging Face releases Carbon-X, a 14B-parameter open model that answers natural-language questions about DNA

Hugging Face released Carbon-X on Thursday, a 14-billion-parameter open model that lets users ask natural-language questions about DNA and get conversational answers. It is designed to run locally on a laptop, according to Wired.

The model was developed with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy. It is trained on genomic sequences and natural language, so users can ask about the effects of DNA changes on gene activity and health.

Key specifications

• Parameters: 14 billion
• Training data: 300 billion tokens combining DNA sequences, English text, and examples linking DNA to biological descriptions
• DNA tokenization: vocabulary expanded to read DNA in groups of six letters
• Input length: trained to ingest sequences of roughly 50,000 base pairs
• Base model: an existing language model, which the team extended (the model is not named in the report)
• Deployment: runs on local hardware; Lewis Tunstall, a senior research scientist at Hugging Face, says no data is sent to Hugging Face or any other company
• Pricing, context window in tokens, and benchmark scores: not yet disclosed

Training data

The DNA sequences come from mammals (including humans), other vertebrates, invertebrates, plants, fungi, and protozoa. Tunstall says the team excluded viruses because they did not want the model used for potentially dangerous applications.

Limits

Users cannot upload full human genomes. Because the model was trained on sequences of roughly 50,000 base pairs, only small parts of a genome can be queried at a time. A human genome contains 6 billion letters.

The developers also caution that they have not validated Carbon-X with wet lab experiments, so they cannot say how accurate its predictions are.

Intended use

Tunstall says the tool is meant to help researchers design new drugs and genetic therapies, and could potentially help patients interpret their own genomes. One use case is judging whether a genetic mutation is benign or disease-causing, a key step in developing a targeted therapy. An intervention aimed at the wrong mutation could be ineffective or cause unintended harm.

"We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA," Tunstall says. He adds that natural language means nontechnical users "can actually ask questions about DNA that wouldn't be accessible through any other tool."

How it differs from other DNA models

Wired contrasts Carbon-X with two existing systems. Google DeepMind's AlphaGenome, released last year, predicts the effects of DNA sequences and mutations in a numeric format and requires more compute. Evo 2, developed by the Arc Institute and Nvidia with academic partners, speaks the language of DNA and can generate entire genomes from scratch. Carbon-X combines DNA and natural language in a conversational interface.

Tunstall says the team deliberately kept the model small. Researchers face restrictions on uploading identifiable or sensitive patient data to frontier AI models, and a 14-billion-parameter model can run on the user's own hardware instead of in the cloud.

What this means

Carbon-X's main contribution is its interface and deployment profile, not a proven accuracy gain. Local execution addresses the data-handling restrictions Tunstall describes, since sensitive sequences stay on the user's machine. The 50,000 base pair input limit confines it to targeted queries rather than whole-genome analysis. With no wet lab validation, and no published benchmark scores, its predictions should be treated as unverified. Excluding viral sequences from training is a deliberate safety choice by the developers, and it also narrows the organisms the model has seen.