Carbon-X: AI model lets users ask questions about DNA sequences in English

On Thursday, 22nd, Hugging Face unveiled Carbon-X, an AI model that takes genetic sequences as input and answers questions posed in English. The aim is to make genetic analysis more accessible through a conversational interface, enabling users to explore the potential effects of mutations and the characteristics of genome segments. The announcement was made by researcher Loubna Ben Allal, a member of the company’s team.

The tool allows users to compare an original sequence with a mutated one and ask about the likely consequences of the change. The response provided is a computational prediction, which may contain errors and requires scientific validation, and does not constitute a medical diagnosis.

How AI analyzes genetic sequences

Carbon-X is built on the Qwen3-14B model and combines DNA and text within a single architecture. This approach enables the correlation of patterns in sequences with information expressed in natural language, rather than relying solely on biological knowledge extracted from documents.

In five tests released by the developers, the average accuracy reached 66.7%, a result close to that of GPT-6.1 Sol, which achieved 67.5%. Two of the tests were designed by the authors themselves. The results assess specific tasks and do not suggest that the tool can diagnose diseases or replace specialists.

In 2025, researchers from InstaDeep had already presented ChatNT in the journal Nature Machine Intelligence, a system that also combines biological sequences with natural language. At that time, a component dedicated to DNA, RNA, and proteins was connected to a language model, foreshadowing the use of a conversational interface for genetic analysis.

Local use requires attention to limitations

The model is available on the Hugging Face Hub under an Apache 2.0 license, including GGUF versions for execution on proprietary infrastructure. The license does not require a mandatory subscription, but operation demands computational resources. The recommended version requires 15.8 GB of storage space and additional memory, as well as software compatible with DNA processing.

Running Carbon-X locally ensures the privacy of genetic data, whereas sending information to a demo or third-party service follows different rules. Developers caution that the tool may invent sequence functions, make positioning errors, and lose quality when generating long segments. Therefore, its use for medical decisions is not recommended.

Frequently Asked Questions

• What is Carbon-X? An AI model that takes genetic sequences as input and answers questions posed in English.
• Can Carbon-X diagnose diseases? It has not been clinically validated and should not be used to guide medical decisions.
• How do I ask Carbon-X questions about DNA? Provide a genetic sequence and a question in natural language; answers are generated in English.
• Do I need to pay a subscription to use Carbon-X? The distribution uses the Apache 2.0 license, which does not require a mandatory subscription; computing costs may apply.
• Is it possible to run Carbon-X on my own computer? Yes, there are GGUF files for local execution, but it is necessary to check available memory and software compatibility.
• Is genetic data kept private in Carbon-X? Privacy depends on the method of access: local execution differs from sending data to third-party services.

The launch is part of the growing use of artificial intelligence in biological research, which includes studies on DNA, gene activity, and initiatives such as Anthropic’s search for enzymes with the help of Claude. In all these cases, it is crucial to distinguish the scope of scientific evidence from promises of clinical application.