Carbon-X: an open model you can chat with about DNA

What is it?

On October 22, Hugging Face's bio team released Carbon-X (Hugging Face, WIRED). It is a 14.8B-parameter decoder-only model trained jointly on DNA sequences and natural-language text. You give it raw DNA in an ordinary chat message and it answers in English.

It predicts whether mutations are likely benign or pathogenic, identifies regions that control gene activity, classifies sequences by organism group, predicts the next bases and explains its reasoning.

How it is built

• Based on Qwen3-14B with no separate DNA encoder: text and DNA share one token stream.
• Hybrid tokenizer: the Qwen3 BPE vocabulary plus 5,515 DNA tokens. DNA is split into non-overlapping 6-mers, so one token covers 6 base pairs.
• 302B tokens of mid-training on DNA, biology text, English and instruction data, followed by SFT and RL with verifiable rewards.
• 40,960-token context; answers come without extended thinking.

How good is it?

Hugging Face reports an average of 66.7% across five DNA benchmarks covering variant effects, sequence understanding and next-base prediction. That is 27 points above the base Qwen3-14B and ahead of every other open model it evaluated. GPT-6.1 Sol scored 67.5%; Carbon-X gets close at roughly 1/1,800 of the inference cost per problem.

The team estimates that analysing the 2.4 million variants of uncertain significance in ClinVar would cost about $300 with Carbon-X, versus $32,000 to $4 million with the frontier models it tested.

How to run it

The bfloat16 weights fit on a single 80 GB GPU, it serves with vLLM and there is a GGUF version. The tokenizer needs trust_remote_code=True and Transformers v5 or later. That lets researchers analyse sensitive or unpublished sequences on their own hardware.

An important limit

The model card is clear: Carbon-X has not been clinically validated and must not be used for medical decisions. As a research tool, though, it is a cheap and open entry point into genomic LLMs.