Hugging Face Bio has released Carbon, a family of open-source generative models designed to predict how genetic variations affect gene activity and health. Carbon-3B, a three-billion-parameter model released in late May 2026, is joined by Carbon-8B and Carbon-500M, while Carbon-A, with 1.2 billion parameters and released on October 8, 2026, directly predicts protein-coding regions. Trained on approximately one trillion tokens drawn from about six trillion DNA and RNA base pairs across eukaryotes, the models deliver up to 275 times the throughput of competing model Evo2-7B on some tasks. The Carbon Annotation Database contains approximately 566 million predicted protein-coding loci, about 16 times the size of RefSeq, and experiments have validated 239 previously unannotated gene candidates in species including cats and chickens. All models and their tools are available under the Apache 2.0 license, but the speed advantage applies only to certain tasks, and the proportion of the 566 million predictions that has yet to be experimentally validated remains unknown.
Source: Read the original article
