Hugging Face released Carbon-X, which CEO and co-founder Clem Delangue calls "the best open-source DNA model," under the Apache 2.0 license. It was made by mid-training Qwen3-14B on 300B tokens of DNA, English and interleaved data, and outperforms all the open language models Hugging Face evaluated on five benchmarks of DNA understanding and generation, a 27-point improvement over Qwen3-14B. Delangue announced it on LinkedIn.
The argument for opening it
Delangue's case is that frontier labs are advancing AI for biology but keeping "the good guys" (scientists, academia, non-profits, health agencies) from using, studying and understanding it, with "widespread but shallow guardrails that the bad guys are bypassing easily." As an example, he says a frontier model blocked him from proofreading his own tweet. He argues this creates "massive asymmetry and risks in the coming years, just like in cybersecurity," especially now that frontier labs are building wet labs, and asks whether this is a cybersecurity scenario for biology. Releasing the model openly, he says, means anyone can study its capabilities, limitations and risks: "Let's make the AI biology frontier open-source to make it safe and impactful for the world." Gated biology access is a theme this blog has followed, for example Anthropic's life sciences verification program.
What Carbon-X is
Carbon-X is a 14.8B-parameter decoder-only model that starts from Qwen3-14B and adds DNA tokens to its vocabulary. A hybrid tokenizer handles English with byte-pair encoding and DNA with a 6-mer tokenizer. Coverage of the model card adds that training ran in four stages, including supervised fine-tuning and reinforcement learning with verifiable rewards, that the weights fit on a single 80 GB GPU, and that Hugging Face worked with the Telethon Institute of Genetics and Medicine, the University of Naples Federico II and Zhongguancun Academy. The point is natural-language questions about DNA sequences, not only sequence prediction.
The numbers
Across the five DNA benchmarks Carbon-X averages 66.7%, against 67.5% for GPT-6.1 Sol and about 39.7% for Qwen3-14B (66.7 minus the 27-point gain). So it comes within 0.8 points of the frontier API model, and Delangue says at 1,800x less cost. All figures are Hugging Face's own, not independently tested, and the model card says it has not been clinically validated and should not guide medical decisions. For other work on DNA models, see DeepMind's AlphaGenome Atlas. Hugging Face is also the subject of NVIDIA's $12.9 billion acquisition deal, covered earlier.