Hugging Face and collaborating researchers released Carbon-X, a 14B model that handles both DNA sequences and natural language. They expanded an existing language model’s vocabulary with six-letter DNA units and trained it on 300 billion tokens combining biological sequences and descriptions. Input is limited to about 50,000 base pairs, so it does not analyze an entire human genome at once. It can run on the user’s hardware, avoiding transmission of sensitive data to external systems. The developers said viruses were excluded from training. Its intended uses are variant interpretation and generation of research hypotheses. Actual laboratory validation has not yet been performed, so diagnostic accuracy cannot be established.
Back to Tech
Carbon-X releases an open model for interpreting DNA through conversation
Local execution can reduce external transfer of genetic data, but predictions without laboratory validation must not be used as diagnoses.