Hugging Face Ships Carbon-X, a 14.8B Open Model That Reads DNA in a Chat Window

Hugging Face released Carbon-X on Thursday, a 14.8-billion-parameter open model that lets a researcher paste raw DNA into an ordinary chat message and get an answer in plain English. The model is published under Apache 2.0, fits on a single 80-gigabyte GPU in bfloat16 and serves through vLLM, which means a genomics lab can run it on hardware it already owns instead of sending patient sequences to a cloud API.

The architecture is deliberately unglamorous. Carbon-X starts from Qwen3-14B and keeps its text vocabulary, then adds 5,515 DNA tokens: 4,096 six-letter combinations known as 6-mers, 1,364 tail tokens for sequences whose length is not a multiple of six, and markers delimiting a DNA span. One DNA token therefore covers six base pairs, and any run of at least six A, C, G or T bases in a prompt is detected and tagged automatically. The model was trained on 300 billion tokens mixing genomic sequence with English text and examples that link the two, drawn from mammals including humans, other vertebrates, invertebrates, plants, fungi and protozoa. Viral genomes were excluded on purpose to reduce misuse risk, and a 6-mer containing any non-canonical base collapses into a single out-of-vocabulary token, so inputs have to be filtered to A, C, G and T first.

The headline numbers come from the model card and have not been independently reproduced. Across five DNA benchmarks Carbon-X averages 66.7 percent with reasoning turned off, within a point of GPT-6.1 Sol at 67.5 percent, at roughly 1/1,800 of the inference cost per problem. Those are vendor figures produced by the team that built the model, and the comparison to a frontier general model is exactly the kind of claim that tends to narrow once outside groups run their own prompts.

The intended task is variant triage, the step that decides whether a mutation is benign or disease-causing before a drug or gene therapy program commits years to the wrong target. Hugging Face senior research scientist Lewis Tunstall said the goal is to help researchers designing drugs and genetic therapies, and eventually to help people interpret their own genomes, adding that the team has unlocked a new capability in language models, the ability to chat with DNA. Because the model runs locally, genomic data never reaches Hugging Face servers, which matters for institutions barred from uploading identifiable patient sequences to third-party services.

The work was done with Zhongguancun Academy in China and the Telethon Institute of Genetics and Medicine with the University of Naples Federico II in Italy. It is the second Carbon release in under two weeks, following Carbon-A, a smaller gene-finding model, and it lands two months after Nvidia agreed to acquire Hugging Face for close to $13 billion in a deal that valued the company as neutral AI infrastructure. A model that runs on a laptop in a lab is a different kind of asset: less of a platform play, more of a demonstration that the company can build domain-specific science tools.

The caveats are substantial and the team states them plainly. There has been no wet-lab validation, so nobody can say how often Carbon-X is right on real biological data. Inputs much longer than those seen in training have not been evaluated, and a full human genome cannot be queried in one pass; the practical unit is a gene and its regulatory region rather than a chromosome. Earlier genomic models such as DeepMind AlphaGenome predict mutation effects but return numeric scores and demand significant compute, while Arc Institute and Nvidia Evo 2 generates sequences at a different scale. Carbon-X sits in front of them as a conversational layer, which is useful for triage and dangerous if mistaken for evidence.

The strategic read is that open, domain-tuned models are becoming the fastest route to scientific adoption. A lab that would never sign a six-figure licensing agreement for a proprietary genomics model can download weights this afternoon, and the resulting dependency runs toward Hugging Face rather than a biotech vendor. But a chat interface that answers confidently about a mutation is exactly the product that needs independent benchmarks before clinicians touch it, and until a wet lab publishes a comparison, Carbon-X is a promising research tool with a very good demo and an unproven error rate.