Hugging Face releases Carbon-X, an open 14.8B model that answers questions about DNA in natural language

Carbon-X lets researchers query genomic sequences in plain language and ask questions about function, variants or regulatory elements without specialized bioinformatics pipelines. Hugging Face built it on Qwen3-14B and trained it on genomic data spanning human, mammalian, plant and fungal genomes, giving it reach beyond the human-centric focus of many bio models.

The headline result is efficiency. Across five DNA benchmarks, Carbon-X averages 66.7%, versus 67.5% for OpenAI's GPT-6.1 Sol, at roughly 1/1,000th the cost. It fits on a single 80GB GPU, which puts it within reach of university labs that cannot send sensitive genomic data to external APIs.

The release echoes a community theme this week. A Hacker News thread (459 points) showed a 10-year-old CNN beating frontier LLMs at skin cancer diagnosis, and specialized, smaller models keep matching or beating generalists in narrow domains. Carbon-X adds evidence that domain post-training on open bases can close most of the gap.

Hugging Face is meanwhile riding broader momentum: its Microduck robot has passed 20,000 preorders at $399, and CEO Clem Delangue is hiring across Pollen Robotics. The caveats for Carbon-X are benchmark breadth and validation. Five benchmarks are a narrow lens, and genomic interpretation errors carry real stakes, so wet-lab and clinical validation will determine whether it moves beyond exploratory research use.