Large language models learned to write emails, debug code and summarize meetings. Hugging Face would now like one to read your genome.
Hugging Face Bio has released a family of open-source generative DNA foundation models called Carbon. They are built to help researchers work out how changes in a genetic sequence might affect gene activity and, ultimately, health.
What Carbon actually does
Think of DNA as an extremely long book written in a four-letter alphabet. Carbon is a model trained to read that book, predict what comes next, and flag which typos actually change the plot.
The flagship is Carbon-3B, released in late May 2026. It sits alongside two siblings, Carbon-8B and Carbon-500M, which Hugging Face Bio describes as intended for speculative decoding, a setup aimed at making generation more efficient.
The newest addition is Carbon-A, a model with 1.2 billion parameters released on October 8, 2026. Its job is narrower and arguably more practical: predicting protein-coding regions directly from DNA across a wide spread of eukaryotic organisms.
Eukaryotes, for the non-biologists, are organisms whose cells keep their DNA in a nucleus. That covers plants, fungi, animals and you.
The models were trained on approximately 1 trillion tokens. That works out to about 6 trillion DNA base pairs, drawn specifically from eukaryotic DNA and RNA sequences.
Out of the box, Carbon handles several core genomics tasks. These include generating sequences, predicting the effects of genetic variants, and analyzing motifs, the short recurring patterns in DNA that often signal biological function.
Smaller model, bigger throughput
Carbon-3B has fewer parameters than rival model Evo2-7B, yet it matches or beats that model on several benchmarks.
It also reportedly delivers up to 275 times faster throughput than Evo2-7B in some tasks.
Part of the trick is how Carbon breaks DNA into pieces. It uses a hybrid tokenizer that combines 6-mers, fixed six-letter chunks of sequence, with Byte Pair Encoding, the same family of technique many text models use to split words into reusable fragments.
That design lets the models handle long stretches of genetic code. Context windows reach up to roughly 393 kbp, or about 393,000 base pairs, in a single pass.
A database 16 times bigger than the reference
Carbon-A’s output may end up being the most immediately useful piece of this release. Hugging Face Bio used it to build the Carbon Annotation Database, which holds approximately 566 million predicted protein-coding loci.
For comparison, that is roughly 16 times larger than RefSeq annotations, a widely used reference resource for genomic data.
Wet-lab validation supported 239 previously unannotated gene candidates in species including cats and chickens. These are genes that standard references had missed, which the model flagged and experiments then backed up.
Why open weights matter here
Every Carbon model and its associated tooling is available on the Hugging Face Hub under the Apache-2.0 license. That is a permissive license that generally allows commercial use, modification and redistribution.
What this means
For researchers, the most direct implication is speed and scale. If Carbon-3B delivers comparable accuracy to larger models at a fraction of the compute, more labs can afford to run variant-effect analysis across large datasets rather than rationing GPU time.
There are caveats worth watching. The 275-fold speed advantage applies in some tasks, not all of them, and benchmark leads often narrow once independent groups run their own comparisons.
The 239 validated gene candidates are also a small slice of a database containing hundreds of millions of predictions. How much of that larger set holds up under experimental testing will determine whether the Carbon Annotation Database becomes a trusted reference or a long list of leads.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.