Hugging Face released Carbon-X on Thursday, an open model that reads DNA sequences and answers questions about them in plain language. It has 14B parameters, an Apache 2.0 licence and runs locally on a laptop, so genetic data never leaves the user's machine. It was built with Zhongguancun Academy in China and the University of Naples Federico II.
Key takeaways
- 14B parameters, base model Qwen3-14B, Apache 2.0 licence
- Trained on 300B tokens mixing DNA, English text and biological descriptions
- Input window is roughly 50,000 base pairs — a full genome cannot be uploaded
- Viral sequences were deliberately excluded from training
- The model reads DNA in six-letter groups
Four parameters that decide where Carbon-X can run:
Why a small model
The size is not a saving but a legal requirement. Researchers face restrictions on sending patient data to frontier models, so the team deliberately built one that fits on a user's hardware.
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.
The team took an existing language model, expanded its vocabulary to read DNA in six-letter groups, and continued training on 300B tokens mixing sequences with text and biological descriptions. The data came from mammals, vertebrates, invertebrates, plants, fungi and protozoa.
AlphaGenome and Evo 2 did it differently
Google DeepMind's AlphaGenome predicts the effects of sequences and mutations, but returns numbers and needs more compute. Evo 2, from the Arc Institute and NVIDIA, speaks the language of DNA and can generate whole genomes. Carbon-X combines DNA and natural language in one conversational interface.
The practical use is deciding whether a mutation is benign or disease-causing — a necessary step in designing a targeted therapy.
| Model | What it returns | Interface |
|---|---|---|
| Carbon-X | an answer in natural language | conversation about a sequence |
| AlphaGenome | numeric values | needs more compute |
| Evo 2 | DNA sequences, up to whole genomes | the language of DNA |
Why it matters
The strongest element is not accuracy, which nobody knows yet, but where the computation happens. A model that fits on a laptop changes the legal risk category for genetic data and opens it to teams that cannot send it to the cloud. The language interface also lowers the barrier for people without bioinformatics training. The open question is whether conversational ease tempts users to treat answers as test results, while nobody has checked them in a lab.
What's next
- The developers have not validated the model in wet lab experiments, so prediction accuracy remains unknown
- NVIDIA agreed in September to acquire Hugging Face for nearly $13 billion
Sources
- WIRED — Hugging Face's New AI Biology Model Lets You Chat With DNA
- Hugging Face — HuggingFaceBio/Carbon-X
