On Oct. 22, 2026, Hugging Face and research partners released Carbon-X, an open language model built to answer questions about DNA in English. The partners named on its model card are Zhongguancun Academy and the Telethon Institute of Genetics and Medicine/University of Naples Federico II. The card lists 14.8 billion parameters and says users can paste raw DNA into a chat to ask about mutation effects or sequence features.
Carbon-X starts from Qwen3-14B. Its developers say they trained the full model on 302 billion tokens: 55% raw DNA, 20% English, 15% biology text and 10% instruction data. A custom tokenizer handles DNA and text in one model, grouping DNA bases into six-base units.
Promising scores, with limits
Across five DNA benchmarks, the developers report a 66.7% average for Carbon-X with extended reasoning switched off, compared with 67.5% for GPT-6.1 Sol at its highest listed reasoning effort. Carbon-X scored ahead of the other open models they evaluated. The developers built two of the five benchmarks, and AiLookout’s Kristian Kostov notes that the results had not been independently reproduced.
The model card also estimates Carbon-X’s inference cost at about 1/1,800 of GPT-6.1 Sol’s cost per problem. That comparison uses measured throughput on an H200 GPU priced at $5 an hour for Carbon-X and standard API pricing for the comparison models.
The Apache 2.0-licensed model card provides download and setup instructions for Transformers and vLLM. The developers say the weights fit on one 80 GB GPU in bfloat16, but warn that llama.cpp and other GGUF runtimes cannot load the custom DNA tokenizer correctly. Running the model locally can keep sequences on a researcher’s own infrastructure; the card cautions that other deployments may handle sensitive genetic data differently.
An introductory post invites users to generate new DNA, but the model card says long-sequence generation degrades and Carbon-X was not trained to design new sequences. It can also miscount bases, confuse coordinates or invent a sequence’s identity or function without enough context. The developers warn that Carbon-X has not been clinically validated and should not be used to make medical decisions.
Comments