Hugging Face Releases Open Carbon-X Model for DNA Processing

Hugging Face, in collaboration with Zhongguancun Academy and two Italian scientific institutes, released Carbon-X, an open (Apache 2.0) 14.8B parameter model fine-tuned from Qwen3-14B for DNA processing: it scores an average of 66.7% on five DNA benchmarks compared to 67.5% for GPT-6.1 Sol, with a claimed inference cost approximately 1,800 times lower than frontier APIs.

Compiled by Sergey KostenchukPublished 2026-10-22Updated 2026-10-22

2026-10-22 Research HuggingFace

🤖 Hugging Face Releases Open LLM for DNA Processing

Carbon-X is an open (Apache 2.0) model with 14.8B parameters, fine-tuned from Qwen3-14B: it reads raw DNA sequences and plain text in a single token stream and fits on a single 80 GB GPU.

🌍 A narrow, fine-tuned open model has nearly caught up to frontier models on DNA tasks at a radically lower inference price—approximately 1,800 times cheaper than frontier APIs. Laboratories will be able to process confidential genomic data locally, without external APIs.

👤 The model can be downloaded from Hugging Face and run locally (vLLM or transformers v5+) to ask questions about DNA variants in a standard chat. Caveat: no clinical validation has been performed; on five DNA benchmarks, it scores 66.7% compared to 67.5% for GPT-6.1 Sol (+27 points over the base Qwen3-14B).

Source 1: https://huggingface.co/HuggingFaceBio/Carbon-X Source 2: https://www.aichatdaily.com/ai-models/hugging-face-releases-carbon-x-open-model-answers-questions

Sources