The model uses a 98,304-base-pair window and predicts on both strands for mammals, other vertebrates, invertebrates, plants, fungi, and protists. Across 42 benchmark genomes, Hugging Face reports a macro-averaged nucleotide F1 of 0.944, ahead of the baselines they ran: AUGUSTUS, Helixer, Tiberius, ANNEVO, OrionGeno, SegmentNT, and NTv3. Hugging Face says the database covers about 11 times as many taxa and 9 times as much sequence as the RefSeq-derived training corpus, about 27 trillion base pairs. A confidence score on each gene reached an AUROC of 0.876 for separating exact coding-sequence matches from other predictions, in their benchmark. Those figures are Hugging Face’s.
The authors note that RefSeq labels are imperfect and change over time. They report that an animal-only checkpoint still annotated plant genomes it had not trained on, and that on Tetrahymena thermophila, which uses a nonstandard genetic code, nucleotide F1 stayed at 0.960. On the pig, the model scored better against a 2026 assembly than against the older 2017 assembly it had seen in training, which the authors read as evidence it learned coding regularities rather than memorizing one annotation.
Initial wet-lab work with ActiveSite and UCSD used PacBio Iso-Seq on cat, Syrian hamster, chicken, and Arabidopsis. Iso-Seq supported some coding regions Carbon-A predicted that are absent from RefSeq. Hugging Face is explicit that this is evidence of transcription, not yet of translation into protein. Ribosome profiling and mass spectrometry are the stated next tests. Hugging Face says EMBL-EBI expects to carry the annotations, and that additional wet-lab results exist and are not public yet.