Hugging Face Carbon-A: 566 Million Candidate Genes Across 22,617 Species, Explained

Hugging Face Bio announced Carbon-A on October 8, 2026: an open model that finds genes in raw DNA, released together with the Carbon Annotation Database, which the team says holds about 566 million candidate genes across 22,617 species. The headline number is easy to misread. These are predictions to be checked, not 566 million newly discovered genes, and the difference is the most useful thing to understand about this release.

This post separates what the announcement says, what the model cards confirm, and what is still unverified. Our sources are the team's own posts on X (Georgia Channing, Thomas Wolf), the Carbon-3B model card on Hugging Face, and the Carbon technical report linked from it.

TL;DR: Carbon-A and Carbon-3B at a glance

What does Carbon-A actually do?

A globe surrounded by books, standing for a Carbon-A inventory of genes across thousands of species

Think of the pipeline in two stages. The first stage reads the sequence and scores every position for how likely it is to sit inside a protein-coding region. The second stage stitches those scores into whole gene models, deciding where exons join and where a gene begins and ends. Splitting the problem this way keeps the large encoder focused on pattern recognition while a lighter model handles the structural bookkeeping that older gene finders encoded as hand-written rules.

For readers who build with AI rather than biology, the lesson is familiar: a general sequence model plus a small task-specific head, trained on curated references, then run at scale to produce a dataset others can build on. The same recipe powers many of the open releases we track, and the release shows how far it can go when the output is a shared public resource instead of a private benchmark score.

Genome annotation is the step after sequencing. A sequencer gives you a very long string of A, C, G and T. Annotation answers: where do the genes start and stop, and which stretches code for proteins? For well-studied organisms this is done with a mix of computational gene finders, RNA data and manual curation. For the thousands of species sequenced in recent years, annotation lags far behind assembly.

According to the team, Carbon-A is an encoder with bi-directional attention that classifies each nucleotide in a eukaryotic genome as coding sequence (CDS) or non-coding. A smaller post-processing model then turns those nucleotide-level calls into gene predictions. The creators say it can annotate a new genome without a closely related reference species, which is the hard case for classic tools that lean on homology.

It is important to say what it is not. It does not design DNA, does not predict what a gene does, and does not claim a gene is functional. It outputs candidate loci.

How big is the database, and what does "candidate" mean?

The numbers reported by the team:

• About 566 million candidate protein-coding loci
• 22,617 taxa, ranging from fungi to mammals
• A work in progress, with plans to keep adding species until GenBank is annotated, so these counts will change

"Candidate" is doing real work. A prediction can be a true unannotated gene, a pseudogene, a fragment, or a false positive. The database is best read as an inventory of leads. Many of the 566 million will overlap with genes already known in reference annotations and with near-duplicate genes across related species, so the number of genuinely novel genes is far smaller than the headline. The team does not claim otherwise, but headlines about the release have been quick to drop the word.

How was it validated?

The team reports two kinds of evidence.

1. Benchmark accuracy. A macro-F1 of 0.944 against RefSeq reference genomes. This measures agreement with existing curated annotations, so it shows the model reproduces known genes, not that its novel calls are right.
2. RNA evidence. Working with Active Site and UC San Diego, the team gathered RNA evidence for 239 genes that are missing from reference annotations in cats, chickens, hamsters and a lab plant. Wet-lab validation was reported for some new genes in well-studied species, including cats, chickens and Arabidopsis. The team says more extensive results on previously unstudied genomes are coming.

That is a reasonable start, and it is also a small fraction of the database. 239 supported genes tells you the approach finds real misses in genomes where we can check; it does not tell you the hit rate in an obscure fungus nobody has studied.

How does it relate to the Carbon DNA language models?

The Carbon family is a separate release from the same Hugging Face Bio effort. The Carbon-3B model card describes a 3B-parameter, decoder-only autoregressive genomic model trained on DNA and RNA sequences, mainly eukaryotes. Key details from the card:

• Native context of 32,768 6-mer tokens (about 197k base pairs), extendable to about 393 kbp at inference with YaRN
• A hybrid tokenizer: non-overlapping 6-mers for DNA plus the Qwen3 BPE vocabulary, left in place for future joint English and DNA training
• Optional species-type and gene-type metadata tokens for conditional generation
• Apache 2.0 license, a public pretraining corpus, and a technical report
• Siblings: Carbon-8B for more capability, and Carbon-500M, a small model for speculative decoding with the larger ones
• Reported as competitive with Evo2-7B on sequence recovery, four variant-effect benchmarks (ClinVar coding and non-coding, BRCA2, TraitGym Mendelian) and two perturbation tasks, while running several times faster, with over 100,000 base pairs per second on a single H100

Secondary coverage (RuntimeWire) repeats a higher figure of roughly 275x throughput over Evo2 and a genome in under two days on one GPU. Those are the team's claims as relayed by a third party; the model card's own wording is "several times faster", so treat the larger multiple as unverified until you read the tech report's hardware and baseline setup.

If you mix the two up, the numbers get confusing. The 566 million figure belongs to the Carbon-A annotation database. The speed and win-rate claims belong to the Carbon language models.

Why this matters

Hugging Face has spent the past year pushing into science, and this is the clearest sign yet. We covered related moves in Hugging Face's ML Intern experiment, the RL environments hub, and the Biohub virtual biology initiative. The pattern: open weights, open data, and a pitch that bio should not be locked inside a few labs.

Three practical reasons it matters:

• Open annotation at scale. Annotating genomes has been a bottleneck for ecology, agriculture and conservation. A freely downloadable first pass lets small groups start from candidates instead of from zero.
• A template for AI-for-science releases. Model, database, tech report and demo ship together, with caveats stated. Compare the plant-focused Gemma 4 Botanic-1 variant ranking and the broader DeepMind survey on AI in science.
• Scientific claims need receipts. The same discipline we apply to lab announcements, such as the AlphaProof Nexus Erdos results, applies here: check the validation, not the headline.

What to check before you use the database

1. Filter by evidence. If a candidate has no supporting RNA or protein evidence for your species, treat it as a lead.
2. Compare with reference annotations. Overlap with RefSeq tells you about agreement; the interesting set is the disagreement.
3. Check the species. Accuracy was benchmarked on well-studied genomes. Distant clades may behave differently.
4. Pin the version. The database is explicitly a work in progress and counts will change.
5. Read the license on each artifact. Carbon-3B is Apache 2.0 per its card; verify Carbon-A and the database separately.

What is claimed vs. what is verified

Related reading on explainx.ai: Hugging Face ML Intern, Gemma 4 Botanic-1, Biohub virtual biology, AI in science survey, AlphaProof Nexus, and Hub RL environments.

Bottom line

Carbon-A is a credible, open step toward annotating the thousands of genomes that have been sequenced but never properly read. The 566 million figure is an inventory of predictions, and the honest summary is: promising accuracy on known genes, early evidence for real misses, and a lot of validation still to do. If you work with non-model organisms, the database is worth downloading as a source of leads. If you only read headlines, remember the word "candidate".

Details reflect the announcement and model cards as of October 13, 2026; the database is actively growing and numbers may change.