Hugging Face releases Carbon-X, an open model for chatting about DNA

Hugging Face's biology team released Carbon-X, a 14.8 billion parameter open language model that can take raw DNA sequences inside a normal chat conversation and respond in English. It can predict whether a mutation is likely harmless or disease-causing, identify regions that control gene activity, classify sequences by organism, predict upcoming bases in a sequence, and explain its reasoning, all from a single model with no separate DNA encoder.

The model is built from Qwen3-14B and uses a hybrid tokenizer: ordinary text uses the standard vocabulary, while DNA is broken into non-overlapping six-base-pair chunks. Hugging Face reports Carbon-X averages 66.7% across five DNA benchmarks, 27 points higher than the Qwen3-14B base model it started from, and close to OpenAI's GPT-6.1 Sol score of 67.5% while costing roughly 1/1,800th as much per problem, by the company's own estimate.

Hugging Face estimates that analyzing the 2.4 million variants of uncertain significance in the ClinVar genetic database would cost about $300 using Carbon-X, compared with $32,000 to $4 million using frontier general-purpose models. The weights fit on a single 80GB GPU and are released under the Apache 2.0 license, letting researchers keep sensitive genetic data on their own infrastructure.

Why it matters

It makes genomic analysis dramatically cheaper and keeps sensitive DNA data off third-party servers, lowering the barrier for biotech research.

Source: Hugging Face