Hugging Face released Carbon-X on Thursday, an open model that reads raw DNA sequences in a chat prompt and answers questions about them in plain English. The model has 14.8 billion parameters, ships under the Apache 2.0 license, and was built with Zhongguancun Academy in China and the Telethon Institute of Genetics and Medicine with the University of Naples Federico II in Italy.
Researchers can paste in a sequence and ask what it does, what a mutation changes, or have the model generate new DNA. The Carbon-X model card lists predicting whether variants are benign or pathogenic, classifying sequences, predicting the next DNA bases, and explaining its reasoning as core tasks. It also handles general instruction-following, scoring 79.3% prompt-level strict accuracy on IFEval.
Carbon-X starts from Qwen3-14B with DNA tokens added to the vocabulary. DNA and English share a single token stream with no separate DNA encoder: text uses standard BPE tokens, and DNA is split into 6-mers. Hugging Face senior research scientist Lewis Tunstall said that scheme encodes DNA with about 3x fewer tokens than BPE.
Key facts
- 01Carbon-X has 14.8B parameters, starts from Qwen3-14B, and is released under the Apache 2.0 license.
- 02It averages 66.7% across five DNA benchmarks versus 67.5% for GPT-6.1 Sol (Max), and 27 points above base Qwen3-14B.
- 03Mid-training used 302B tokens: about 55% raw DNA, 20% English web text, 15% biology text and 10% instruction data.
- 04Built with Zhongguancun Academy and Italy's Telethon Institute of Genetics and Medicine with the University of Naples Federico II.
- 05The model is not clinically validated and its predictions have not been tested in wet-lab experiments.
Mid-training ran on 302 billion tokens, roughly 55% raw DNA, 20% English web text, 15% biology text, and 10% instruction data. Hugging Face followed that with supervised fine-tuning on 20 dataset subsets and reinforcement learning with verifiable rewards on DNA tasks. The DNA came from humans, other mammals, plants, and fungi; viruses were excluded to reduce misuse risk.
On benchmarks with thinking disabled, Carbon-X averages 66.7% across five DNA tests, against 67.5% for OpenAI's GPT-6.1 Sol (Max) and 27 points above the base Qwen3-14B. Individual scores include 85.1 on BioReason Non-SNV, 74.9 on BioReason Coding, 70.4 on FrontierVEP, 63.6 on GenomeQA, and 39.4 on NextBase.
Tunstall said Carbon-X beats every open language model Hugging Face tested on DNA benchmarks and lands within a point of GPT-6.1 Sol at more than 1,000x lower inference cost. Hugging Face CEO Clem Delangue put the cost gap at 1,800x.
Carbon-X runs on researchers' own hardware rather than through a cloud API, which means sensitive patient genomic data does not have to leave the lab. Its context window is 40,960 tokens, and it can analyze about 50,000 base pairs at a time, not full genomes. Tunstall framed local deployment as a way to give scientists independence from a single vendor.
“tools that don't make them fully dependent on some external single provider”— Lewis Tunstall, Senior research scientist at Hugging Face
Hugging Face says the model could help researchers judge whether mutations are benign or disease-causing, which could inform drug and gene-therapy design. Tunstall pointed to individualized treatment as the long-term target.
“the holy grail is, in general, this concept of personalized medicine”— Lewis Tunstall, Senior research scientist at Hugging Face
The company is explicit about limits. Carbon-X is not clinically validated, its predictions have not been tested in wet-lab experiments, and Hugging Face warns against using it for medical decisions. It positions the model as a conversational layer alongside more compute-heavy DNA systems such as DeepMind's AlphaGenome and the genome-generating Evo 2.
Nvidia agreed in September to acquire Hugging Face for nearly $13 billion. Weights are available now on the Hugging Face Hub.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
