Carbon-X

Carbon-X is a 14.8B-parameter decoder-only language model from Hugging Face's bio team (Carbon family), trained jointly on DNA sequences and natural-language text. It was started from Qwen3-14B. DNA and text share one token stream through a hybrid tokenizer that keeps the Qwen3 BPE vocabulary for text and encodes DNA as non-overlapping 6-mers. There is no separate DNA encoder.

Users paste raw DNA into a chat prompt and get English answers. Supported tasks include predicting whether variants are benign or pathogenic, identifying gene-regulatory regions, classifying sequences by organism group, predicting next bases, and explaining its reasoning. Training used about 302B tokens of mid-training on DNA, biology text, English and instruction data, followed by supervised fine-tuning and reinforcement learning with verifiable rewards.

On five DNA benchmarks the model card reports an average of 66.7% with thinking disabled, 27 points above Qwen3-14B and close to a frontier model (67.5%) at much lower inference cost per problem. These figures are vendor-reported. The weights fit on a single 80 GB GPU in bfloat16 and serve with vLLM, so sequences can stay on local infrastructure. It requires Transformers v5 or later and trust_remote_code=True for the tokenizer. The context window is 40,960 tokens, and it is a research model, not a clinical tool.