Hugging Face releases Carbon-X open DNA language model

Bridging DNA and English

Hugging Face Bio has launched Carbon-X, a new open-weights model designed to interpret raw DNA sequences. Unlike previous systems that required separate encoders for different types of data, Carbon-X uses a single decoder to handle both DNA and natural language. This means users can provide raw genetic code as input and receive explanations or analysis in plain English.

The model was developed in collaboration with the Zhongguancun Academy and TIGEM / University of Naples Federico II. It is built on the Qwen3-14B architecture but has been specifically tuned for biological tasks. It can perform several functions, such as predicting whether mutations are benign or pathogenic, flagging regulatory regions, and classifying various organisms.

By releasing the weights under the Apache 2.0 license, Hugging Face ensures that research labs can maintain data privacy. Because the model can run on a single 80 GB GPU, organizations can process sensitive genetic sequences on their own infrastructure. This avoids the need to upload private biological data to external cloud providers for analysis.

Hybrid Tokenization and Training

A core technical feature of Carbon-X is its hybrid tokenizer. It retains the existing vocabulary for English text while adding 5,515 new tokens specifically for DNA. These include 4,096 non-overlapping 6-mers, which are short sequences of six base pairs, along with tail tokens and specific markers like <dna> and </dna>.

The training process occurred in four distinct stages to balance biological knowledge with linguistic fluency. It began with an embedding warmup where the transformer body remained frozen while the model learned new DNA embeddings. This was followed by a mid-training phase using 302B tokens, where the data mix included 55% raw DNA, 20% English, 15% biology text, and 10% instruction data.

The team faced challenges with the learning rate during development. High rates allowed the model to learn DNA quickly but caused it to forget how to speak English. By using a lower learning rate and a 14B parameter size over a long run, the developers successfully balanced these two capabilities. The final stages involved supervised fine-tuning and reinforcement learning using verifiable rewards.

Benchmarking and Performance

Carbon-X shows strong performance across various genomic metrics, achieving a 66.7% average on five benchmarks. In several categories, it outperformed every other open model tested by Hugging Face. Specifically, it led in BioReason Coding tasks with a score of 74.9% and excelled at next-base prediction on the NextBase benchmark.

While it is highly capable, the model still faces limitations compared to the largest frontier models. It trails behind GPT-6.1 Sol on complex variants and single-base variants. Additionally, its performance on specific tasks like splice-site questions was lower, scoring 29.8% on the GenomeQA benchmark. The model card notes that it is not clinically validated and can still hallucinate sequence identities.

The model also maintains its general intelligence despite its biological specialization. It achieved a 79.3% strict accuracy on the IFEval, which measures general instruction following. This suggests that the model remains useful for general-purpose tasks while providing specialized capabilities for DNA research. It can even serve as a sub-agent to improve the performance of other models, such as Kimi K3.

Cost and Deployment

One of the most significant advantages of Carbon-X is its economic efficiency for large-scale research. Hugging Face estimates that scoring 2.4 million ClinVar variants would cost approximately $300 using Carbon-X. In contrast, performing the same task on frontier models could cost between $32,000 and $4 million depending on the specific model and effort level.

The model is designed to be accessible to researchers with standard hardware setups. It can be deployed using vLLM or Transformers on a single 80 GB GPU. For those looking for even lighter requirements, the official GGUF version in Q4_K_M format takes up only 9.6 GB of space. This makes it feasible for smaller labs to run high-quality DNA analysis without massive compute budgets.

For developers, the repository includes several tools to integrate the model into existing workflows. These include an OpenAI-style tool, an Inspect agent, and a carbon-dna skill for coding agents. The team recommends keeping the thinking feature disabled during use, as the model was specifically post-trained and evaluated in that configuration. These tools help bridge the gap between raw model weights and practical research applications.