Hugging Face unveils Carbon, an open-source AI model family for reading DNA

Large language models learned to write emails, debug code and summarize meetings. Hugging Face would now like one to read your genome. Hugging Face Bio has released a family of open-source generative DNA foundation models called Carbon. They are built to help researchers work out how changes in a genetic sequence might affect gene activity and, ultimately, health. What Carbon actually does Think of DNA as an extremely long book written in a four-letter alphabet. Carbon is a model trained to read that book, predict what comes next, and flag which typos actually change the plot. The flagship is Carbon-3B, released in late May 2026. It sits alongside two siblings, Carbon-8B and Carbon-500M, which Hugging Face Bio describes as intended for speculative decoding, a setup aimed at making generation more efficient. The newest addition is Carbon-A, a model with 1.2 billion parameters released on October 8, 2026. Its job is narrower and arguably more practical: predicting protein-coding regions directly from DNA across a wide spread of eukaryotic organisms. Eukaryotes, for the non-biologists, are organisms whose cells keep their DNA in a nucleus. That covers plants, fungi, animals and you. T...