Hugging Face has released Carbon-X, an open, 14.8-billion-parameter model trained on DNA sequences and natural-language text. Its model card says users can put raw DNA into a chat and get English answers about questions such as whether a mutation is likely to be benign or harmful.
The model card describes a tokenizer that reads DNA in non-overlapping groups of six bases while using a separate vocabulary for text. Lewis Tunstall, a senior research scientist at Hugging Face, gave Business Insider the example of a researcher asking whether a mutation could increase disease risk. WIRED reports that Tunstall says the tool is meant to help researchers design new drugs and genetic therapies.
What the tests do and don't show
In its evaluation, Hugging Face Bio reports an average score of 66.7% across five DNA benchmarks with thinking disabled, compared with 67.5% for GPT-6.1 Sol. Those are developer-reported results: the tasks use different scoring measures, the evaluation uses one sample per problem, and the developers built two of the five benchmarks.
The model card warns that Carbon-X has not been clinically validated and should not be used to make medical decisions. It says the model can make mistakes in exact sequence operations or invent a sequence's function when given too little context. WIRED reports that the developers have not validated its predictions in wet-lab experiments.
A laptop claim conflicts with the model card
In a release post, _lewtun claims Carbon-X runs locally on a laptop with llama.cpp. The model card instead says llama.cpp and other GGUF runtimes cannot load its custom Python DNA tokenizer and would tokenize DNA incorrectly. It documents Transformers and vLLM, with about 80 GB of GPU memory specified for one agent setup.
The model card says its weights fit on one 80 GB GPU in bfloat16, allowing researchers to keep sequences on their own infrastructure. It also cautions that data handling differs across local downloads, APIs, third-party providers and demos.
