Hugging Face and academic collaborators have released Carbon-X, an open 14-billion-parameter model that accepts both DNA sequences and ordinary language. Researchers can ask conversational questions about how a genetic change may affect gene activity or health.
The team expanded an existing language model’s vocabulary to read DNA in six-letter groups, then trained it on 300 billion tokens combining genomic sequences, English and biological descriptions. Training data covered mammals, plants, fungi and other organisms, but excluded viruses to reduce the potential for dangerous use.
Carbon-X can process about 50,000 base pairs at a time, not an entire human genome. Its smaller size is intended to let laboratories run it locally, keeping sensitive genomic data off third-party cloud services. Potential uses include judging whether a mutation is likely benign or disease-causing during early drug or gene-therapy research.
The main limitation is evidence: the developers have not yet validated its predictions in wet-lab experiments and do not claim established accuracy for clinical decisions. Carbon-X therefore offers a new interface for genomic analysis, not a diagnostic tool or a substitute for experimental confirmation.