Reading DNA like a chat
Hugging Face Bio has released Carbon-X, a 14.8-billion-parameter model with open weights. It takes raw DNA sequences through natural-language prompts and responds in English. Its tasks include predicting whether mutations are benign or pathogenic, flagging regulatory regions, classifying organisms and predicting the next base.
Technical design
Carbon-X is built on Qwen3-14B and processes DNA and text in a single token stream. There is no separate DNA encoder. A hybrid tokenizer keeps Qwen3's BPE vocabulary for text while adding 5,515 DNA tokens, 4,096 of which correspond to six-base chunks (6-mers). The model has a 40,960-token context window and can run in bfloat16 on a single 80 GB GPU. Official GGUF versions are also available: Q8_0 at 15.8 GB and Q4_K_M at 9.6 GB.
Training and performance
Training ran in four stages: embedding warmup, 302 billion tokens of mid-training, supervised fine-tuning and reinforcement learning with verifiable rewards. According to Hugging Face's self-reported results, the model averages 66.7% across five DNA benchmarks, the best among open models tested. It scores 74.9% on BioReason Coding, ahead of every open and closed model tested. Its weakest result is 29.8% on GenomeQA splice-site questions.
Why it matters
The Apache 2.0 license allows labs to keep sensitive sequences on their own GPUs. The model also retains general language skills, with 79.3% prompt-level strict accuracy on IFEval. However, all scores are self-reported by Hugging Face; independent verification is not yet available.
