Hugging Face lists Carbon-X on its Hub as of Oct. 22, an open 14.8-billion-parameter model that takes raw DNA in a chat conversation and answers in English. The weights are free to download under an open Apache 2.0 license.

Hugging Face says it can judge whether a mutation looks benign or harmful, flag regions that control gene activity, identify what group of organisms a sequence comes from, and predict the next letters of a sequence.

What it is

  • Base: Qwen3-14B, an existing language model, with DNA added to its vocabulary so DNA and text share one stream.
  • Training: 302 billion tokens, about 55% of it DNA. Fine-tuning and reinforcement learning followed.
  • Partners: Hugging Face, Zhongguancun Academy, and the Telethon Institute of Genetics and Medicine with the University of Naples "Federico II".
  • Input: training sequences ran up to about 49,000 DNA letters.

Who can use it

It runs on one 80 GB GPU, so labs can keep unpublished sequences on their own machines. Hugging Face says it works with its Transformers library (version 5 or later) or the vLLM serving tool. Hugging Face says llama.cpp and other GGUF tools are not supported.

Benchmark claims

Hugging Face reports an average of 66.7% across five DNA benchmarks, 27 points above Qwen3-14B. It lists GPT-6.1 Sol at 67.5%, and says Carbon-X costs about 1/1,800 as much per problem. It says Carbon-X leads every model it tested on the BioReason Coding benchmark (74.9) and on next-base prediction (39.4). Hugging Face built two of the five benchmarks itself, FrontierVEP and NextBase.

What it cannot do

Hugging Face says Carbon-X is not clinically validated and should not be used to make medical decisions. It can miscount bases, confuse coordinates and make incorrect edits, and it may invent a sequence's identity or function. It reads and explains DNA; it was not trained to design new sequences.

Why it matters

A 14.8B-parameter model that fits on one 80 GB GPU means, per the model card, that researchers can keep sensitive or unpublished sequences on their own infrastructure. The trade-off is trust. The scores are Hugging Face's own, and the card warns against using it for clinical calls, so it is a research aid, not a diagnostic.

For another open release aimed at specialists, see Microsoft's open-source weather forecast tool. For the other end of the size scale, see Meta's open Muse Spark weights.