Artificial Intelligence

Hugging Face Releases Carbon-X DNA Language Model

By Interestana AI EditorialAI summary of reporting by MarkTechPostHow we report

MarkTechPost2 min read

Hugging Face has released Carbon-X, an open DNA language model with 14.8 billion parameters. Developed by Hugging Face Bio, it is a decoder-only model capable of reading raw DNA and responding in English.

Key points

  • Hugging Face released Carbon-X, a 14.8B parameter open DNA language model.
  • Carbon-X integrates DNA and natural language processing for biological analysis.
  • The model predicts DNA mutations, regulatory regions, and classifies organisms.
  • It achieved 66.7% average performance on five DNA benchmarks.
  • Carbon-X uses a hybrid tokenizer for DNA and text processing.

Carbon-X can predict whether DNA mutations are benign or pathogenic, identify regulatory regions, classify organisms, and predict the next bases in a DNA sequence. Its weights are available under the Apache 2.0 license, allowing labs to process sensitive sequences on their own GPUs.

The model was built in collaboration with Zhongguancun Academy and TIGEM / University of Naples Federico II. It features a 40,960-token context and can run on a single 80 GB GPU using bfloat16 precision via vLLM or Transformers.

By the numbers

Carbon-X parameters
14.8B
token context
40,960
average DNA benchmark performance
66.7%

Carbon-X achieved 66.7% average performance on five DNA benchmarks, reportedly outperforming all other open models tested. It reached a high of 74.9% on the BioReason Coding benchmark, surpassing both open and closed models.

The model is based on Qwen3-14B and utilizes a hybrid tokenizer that combines Qwen3's BPE vocabulary for text with 5,515 DNA tokens. One DNA token represents six base pairs, and DNA runs of six or more bases are automatically tagged.

Training involved four stages: embedding warmup with the transformer frozen, mid-training on a mix of DNA, English, biology text, and instruction data, followed by supervised fine-tuning (SFT), and reinforcement learning (RL) using asynchronous GRPO.

Sources and references

Original report

Hugging Face Releases Carbon-X, a 14B Open DNA Language Model that Brings DNA and Natural Language Together for the First Time

MarkTechPost

Read on MarkTechPost (opens in a new tab)

Ref 00C0ELDW

Share this story

Cite this

APA

Interestana AI Editorial. (2026, October 25). Hugging Face Releases Carbon-X DNA Language Model. Interestana. https://interestana.com/articles/hugging-face-releases-carbon-x-a-14b-open-dna-language-model-that-brings-dna-and-natural-language-together-for-the-first-time-00c0eldw

MLA

“Hugging Face Releases Carbon-X DNA Language Model.” Interestana, 25 Oct. 2026, interestana.com/articles/hugging-face-releases-carbon-x-a-14b-open-dna-language-model-that-brings-dna-and-natural-language-together-for-the-first-time-00c0eldw.

Chicago

Interestana AI Editorial. “Hugging Face Releases Carbon-X DNA Language Model.” Interestana, October 25, 2026. https://interestana.com/articles/hugging-face-releases-carbon-x-a-14b-open-dna-language-model-that-brings-dna-and-natural-language-together-for-the-first-time-00c0eldw.

BibTeX

@misc{interestana_00c0eldw,
  author       = {{Interestana AI Editorial}},
  title        = {{Hugging Face Releases Carbon-X DNA Language Model}},
  howpublished = {Interestana},
  year         = {2026},
  month        = oct,
  url          = {https://interestana.com/articles/hugging-face-releases-carbon-x-a-14b-open-dna-language-model-that-brings-dna-and-natural-language-together-for-the-first-time-00c0eldw},
  note         = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/25/hugging-face-carbon-x-14b-dna-language-model/}
}

Credit the source

Original reporting: MarkTechPost, “Hugging Face Releases Carbon-X, a 14B Open DNA Language Model that Brings DNA and Natural Language Together for the First Time”, https://www.marktechpost.com/2026/10/25/hugging-face-carbon-x-14b-dna-language-model/(opens in a new tab)

Follow this story

Keywords

Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.

Subscribe

Read next