Hugging face introduces Carbon-X, a 14b-parameter open dna language model

Hugging Face has launched Carbon-X, an open-source DNA language model with 14.8 billion parameters, marking the first integration of DNA and natural language processing. The model, developed in collaboration with Zhongguancun Academy, TIGEM, and the University of Naples Federico II, is a decoder-only model from Hugging Face Bio. It processes raw DNA sequences and responds in English, offering capabilities such as mutation classification, regulatory region detection, and next-base prediction. The model’s weights are released under Apache 2.0, allowing labs to retain sensitive data on their own hardware. Carbon-X is built on Qwen3-14B, utilizing a hybrid tokenizer that combines Qwen3’s BPE vocabulary with 5,515 DNA-specific tokens. Training involved four stages, with learning rate adjustments playing a critical role in balancing DNA and English performance. The model outperforms Qwen3-14B by 27 points and trails GPT-6.1 Sol by 0.8 points, excelling in coding-variant effects and next-base prediction. It also maintains strong general skills, achieving 79.3% strict accuracy on IFEval. Hugging Face estimates Carbon-X’s inference cost at $5 per hour on an H200, significantly lower than frontier models.