Carbon-X: the best open DNA model (huggingface.co)

The unveiling of Carbon-X, an open 14.8B-parameter language model specifically designed for DNA comprehension and generation, marks a significant advancement in the AI/ML community. This model is capable of interpreting raw DNA sequences and producing human-readable explanations in English. It can predict the pathogenicity of mutations, classify biological sequences, and even forecast the next bases in a given DNA sequence—all within a single conversational context. This integration of DNA and natural language through a hybrid tokenizer allows for efficient processing without the need for separate encoders. Carbon-X stands out in performance, achieving an impressive average score of 66.7% across five DNA-related benchmarks, significantly outperforming its predecessor, Qwen3-14B, by 27 points. Its architecture features a contextual window of 40,960 tokens and a custom hybrid tokenizer that incorporates both DNA sequences and text. With low operational costs—an estimated $300 to analyze 2.4 million variants compared to $32,000 to $4 million using frontier models—Carbon-X provides researchers a powerful yet accessible tool for genomic analysis, enabling them to retain sensitive data on their infrastructure. The efficiencies realized with Carbon-X could revolutionize practices in genomics, making advanced bioinformatics tools more attainable for researchers worldwide.