Hugging Face released Carbon-X on October 22. It is an open 14.8B-parameter language model that reads DNA and answers in English. The model is built on Qwen3-14B and licensed under Apache 2.0.

What it does

According to the model card, you paste raw DNA into an ordinary chat and ask questions in English. Hugging Face says the model can predict whether a mutation is benign or pathogenic, identify regions that control gene activity, classify sequences by organism group and predict the next bases of a sequence. These are research capabilities. The model card says Carbon-X "has not been clinically validated, and it should not be used to make medical decisions." Hugging Face does not say it can be used for clinical diagnosis, and we have not tested it. The card also notes that quality drops on long DNA generation, and that the model "was not trained to design new" sequences.

How it was trained

Carbon-X starts from Qwen3-14B and adds 5,515 DNA tokens to the vocabulary, encoding DNA as 6-base chunks. The model card describes 302B tokens of mid-training on DNA, biology text, English and instruction data, followed by supervised fine-tuning and reinforcement learning on DNA tasks.

Hugging Face's numbers

All of the following come from Hugging Face's own tests and have not been checked by a third party:

  • Across five DNA benchmarks, Carbon-X averages 66.7%, against 67.5% for GPT-6.1 Sol, at about 1/1,800 of the inference cost per problem. Carbon-X costs are based on measured throughput at $5 per H200-hour. Comparison models use standard API pricing.
  • Hugging Face says that score is 27 points above Qwen3-14B.
  • Using Carbon-X as a DNA sub-agent raised Kimi K3's score on the same five benchmarks from 49.4% to 60.5%, a gain of 11.1 points.

Hardware

The model card says the weights fit on a single 80 GB GPU in bfloat16 and runs with vLLM or Transformers. Hugging Face post-training researcher Lewis Tunstall wrote on X that Carbon-X runs locally on a laptop with llama.cpp, and the release blog shows a demo on a MacBook. The model card's usage section, however, currently states that llama.cpp and other GGUF runtimes are not supported because the DNA tokenizer is custom Python code. For another open model built to run on your own hardware, see MiniCPM-V 4.7.

Sources

  • Hugging Face's Carbon-X model card
  • Hugging Face Space: Carbon-X release blog
  • Loubna Ben Allal's post on X
  • Lewis Tunstall's post on X