Hugging Face's Carbon-X Is an Open Model That Chats About DNA
Hugging Face released Carbon-X, a 14.8-billion-parameter open-weight model, under Apache 2.0. Built from Qwen3-14B and trained on 302 billion mid-training tokens, mostly raw DNA, it answers questions about genome sequences inside ordinary chat and predicts whether variants are benign or pathogenic. The model card reports 66.7 percent across five DNA benchmarks, close to GPT-6.1 Sol at a claimed fraction of the cost.
On this page
What Carbon-X is
Carbon-X is a 14.8-billion-parameter decoder-only model, released under Apache 2.0, that treats DNA and English as one language. There is no separate DNA encoder: the vocabulary of 157,184 tokens combines ordinary BPE text tokens with 5,515 DNA tokens that encode sequences as 6-mers, so you paste raw genetic sequence into a chat conversation and the model answers in English. The release sits inside a broader Carbon family on the HuggingFaceBio organization, alongside Carbon-500M, Carbon-3B and Carbon-8B, GGUF quantizations of each, and a Carbon-A annotation model, with training code under the same license on GitHub.
The work the model card emphasizes is variant effect prediction: given a reference sequence and a variant, Carbon-X answers whether it is likely benign or pathogenic and explains its reasoning. It also identifies regulatory regions such as enhancers and promoters, splice sites and transcription factor binding sites, and classifies sequences by organism group.
How it was trained
Carbon-X starts from Qwen3-14B. A vocabulary-extension phase on roughly 10 billion tokens teaches the base model the DNA tokens with the body frozen, then 302 billion tokens of mid-training follow, a mixture of 55 percent raw DNA, 20 percent English, 15 percent biology text and 10 percent instruction data. Supervised fine-tuning comes next, then two rounds of asynchronous GRPO with verifiable rewards, a recipe that uses checkable answers rather than a reward model.
The headline numbers come from five DNA benchmarks, where Carbon-X averages 66.7 percent, 27 points above its Qwen3-14B base and close to GPT-6.1 Sol at 67.5 percent. The model card puts the cost difference at roughly 1,800 times, estimating about 300 dollars to analyze ClinVar's 2.4 million variants of uncertain significance against 32,000 to 4 million dollars with frontier models.
What it can and cannot do
The context window is 40,960 tokens, which works out to about 49,000 base pairs of DNA, so this is gene-scale analysis, not whole-genome. The card is also candid about limits: Carbon-X was not trained to design new sequences, and long DNA generation degrades. For deployment, the weights fit on a single 80 GB GPU in bfloat16 and the recommended serving path is vLLM with the custom tokenizer, which means GGUF and llama.cpp are not supported for this model, unlike the smaller Carbon siblings whose GGUF builds are already on the hub.
One detail will interest the agent crowd: the card ships an ask_carbon helper and a Claude Code and Codex skill for querying the model from a coding agent, with the practical warning to pass DNA sequences by reference because agents miscopy long strings.
Why an open DNA model matters
Genomic models have been a specialized corner of AI where the interesting artifacts were mostly closed or academic. A family that spans 500 million to 14.8 billion parameters, all Apache 2.0 with GGUF builds for the small ones, changes who can experiment: a university lab with one GPU can now run the same weights that anchor the benchmark comparisons, and the training recipes are public.
For readers of this blog the pattern is familiar. Specialized open models keep proving that a mid-size base model plus domain data can sit near a frontier system on that domain's benchmarks, and the infrastructure story is the same one as always: know the context limits, check what quantization does to your task, and read the model card's own list of what the model was not trained to do.