Carbon-X: an open DNA language model
On 22 October 2026 Hugging Face, Zhongguancun Academy and the Telethon Institute of Genetics and Medicine with the University of Naples Federico II published the weights of Carbon-X, a 14.8-billion-parameter language model trained jointly on DNA sequences and English text, under Apache 2.0, with a quantised GGUF version. A user puts raw DNA into an ordinary chat and the model answers in English, for instance whether a variant is likely benign or pathogenic.
Why it matters
A model that reads DNA and text in one token stream, small enough for a single 80 GB graphics card, is available with open weights, so a lab can keep sensitive or unpublished sequences on its own machines instead of sending them to a frontier model's API. The card says it has not been clinically validated and that it reads existing sequences rather than designing new ones.
What the card says. Carbon-X starts from Qwen3-14B and is trained on 302 billion tokens of mid-training on DNA, biology text, English and instruction data, then by supervised fine-tuning and reinforcement learning with verifiable rewards on DNA tasks. A hybrid tokenizer keeps the text vocabulary and encodes DNA as non-overlapping 6-mers, so one DNA token covers six bases; there is no separate DNA encoder. The context is 40,960 tokens; the model was post-trained and evaluated with thinking disabled. It is offered for predicting whether mutations are benign or pathogenic, for finding regions that control gene activity, for classifying sequences by organism group and for predicting next bases. The GGUF repository needs a llama.cpp build from b11742 or later. Dates. The first commit of the repository's present history is 22 October 2026 at 09:55 UTC (a page rendered for a viewer in the US Pacific zone shows it as 22 October 02:55); the repository itself was created on 21 October. The card has no date of its own, and the blog it links is a Hugging Face Space that was not read. What the record does not claim. The card compares Carbon-X with frontier models on five DNA benchmarks and gives costs per problem; these are the authors' own evaluations and none of their figures is repeated as a finding here. The card lists its own limits: splice-site and histone-mark questions remain weak, the model can miscount bases and confuse coordinates, it may invent a sequence's identity when the prompt lacks context, and quality degrades when it generates long sequences. It also warns that genetic data is highly sensitive.