Hugging Face's open Carbon-X model chats about DNA
Why it matters
Hugging Face released Carbon-X on Oct 22, an open DNA language model built with Zhongguancun Academy and the Telethon Institute of Genetics and Medicine / University of Naples Federico II. You paste raw DNA into an ordinary chat and it answers in English: it predicts whether mutations are benign or pathogenic, identifies regions that control gene activity, predicts the next bases of a sequence and explains its reasoning. The 14.8B-parameter model starts from Qwen3-14B and was mid-trained on 302B tokens of DNA, biology text, English and instruction data, reading DNA in 6-base tokens. Across five DNA benchmarks it averages 66.7% with thinking off, 27 points above Qwen3-14B and ahead of every other open model tested, close to GPT-6.1 Sol's 67.5% at about 1/1,800 of the inference cost per problem. It is the top model tested on coding-variant effects (74.9) and next-base prediction (39.4); frontier models lead the other three benchmarks. Hugging Face estimates analysing ClinVar's 2.4 million variants of uncertain significance would cost about $300 with Carbon-X, versus $32,000 to $4 million with the frontier models tested. The Apache 2.0 weights fit on a single 80 GB GPU in bfloat16, so labs can keep sensitive sequences in-house. Caveats: it is not clinically validated and should not be used for medical decisions; splice-site (29.8%) and histone-mark (34.6%) questions remain weak; and these are Hugging Face's own evaluations, on five benchmarks, two of which it built.
Read the source: Hugging Face (opens in a new tab)
All hits from 23 Oct 2026More Open ModelsGet the daily hits by email