Hugging Face has released Carbon-X, an open 14.8-billion-parameter language model that reads raw DNA inside an ordinary chat conversation and answers in plain English. The model is out under the Apache 2.0 license, so anyone can download it, change it and run it locally, the company said on October 22.
Carbon-X is the latest model in Hugging Face’s Carbon family and a first-of-its-kind combination of DNA analysis and plain language: paste in a genetic sequence, ask whether a mutation looks benign or disease-causing, and the model answers like a chatbot, explaining its reasoning. It was built with researchers at the Zhongguancun Academy in China and the Telethon Institute of Genetics and Medicine at the University of Naples Federico II in Italy.
What Carbon-X is
Carbon-X is a decoder-only language model that starts from Qwen3-14B, Alibaba’s open model, extended with DNA tokens so DNA and text share a single token stream — there is no separate DNA encoder. A hybrid tokenizer keeps the Qwen3 BPE vocabulary for text and encodes DNA as non-overlapping 6-mers, so one DNA token covers six base pairs. The context window is 40,960 tokens.
Training ran to 302 billion tokens — about 55% of them raw DNA — followed by supervised fine-tuning and reinforcement learning with verifiable rewards. The genomes in the corpus came from humans and other mammals, other vertebrates, invertebrates, plants, fungi, protozoa and microbes, plus English and biology text. The team deliberately excluded viruses to prevent potentially dangerous applications.
What it can do
According to the model card, Carbon-X predicts whether mutations are benign or pathogenic, identifies regions that control gene activity, classifies sequences by organism group, predicts the next bases of a sequence and explains its reasoning. Responses can mix English explanations with DNA, and earlier biology models that returned only numbers or raw sequences now have a conversational competitor.
The headline numbers: across five DNA benchmarks covering variant-effect prediction, sequence understanding and next-base prediction, Carbon-X averages 66.7% — 27 points above the Qwen3-14B it starts from and ahead of every other open model Hugging Face evaluated. It lands close to GPT-6.1 Sol at 67.5%, which Hugging Face notes costs about 1,800 times more per problem in inference. It leads every model tested on the BioReason Coding benchmark and on next-base prediction.
Practicality is part of the pitch. The weights fit on a single 80 GB GPU in bfloat16, and the model serves with vLLM — so labs can keep unpublished or sensitive sequences on their own machines instead of uploading them to a cloud service. Lewis Tunstall, a senior research scientist at Hugging Face, said the team designed it to run locally so no genomic data is transmitted to any company.
Why It Matters
Carbon-X is aimed squarely at a bottleneck in modern biology: researchers who need AI help with genomic data but cannot send patient sequences to a frontier lab’s API. A free, open model that runs on one GPU and talks in plain English makes genomic analysis accessible to people who are not trained bioinformaticians — and keeps the data local.
Tunstall said the deeper motivation was to stop biology researchers from becoming dependent on the major AI labs. “If you want to really push the frontiers of mathematics, you likely have to be at one of these AI labs like OpenAI or Anthropic,” he said, suggesting the same concentration was coming for science — and that open models like Carbon-X are the counterweight.
Background
Hugging Face, the model-sharing platform which NVIDIA agreed to acquire for nearly $13 billion in September, has been steadily releasing its own research models alongside the community’s. The Carbon family has explored genomic modeling before, but Carbon-X is the first to fuse DNA and natural language in a single chat interface — treating DNA, as the team puts it, as just another language a chat model can read.
Source: Hugging Face (official model card) ; SiliconReport
0 Comments