Genomics & Single-Cell · Hugging Face, from GENERanno (Gener Team)
Carbon-A
Predicts eukaryotic protein-coding gene structures and per-gene confidence scores directly from genomic DNA.
What Carbon-A is
Carbon-A is a 1.2-billion-parameter genomic foundation model developed by Hugging Face from GENERanno (Gener Team). It identifies protein-coding sequence (CDS) regions at single-base resolution across eukaryotic genomes, generating strand-specific probabilities that preserve coding regions overlapping on opposite strands.
A dedicated annotation pipeline transforms the base-level probabilities into complete gene structures. The pipeline scores boundary sites, applies a grammar-constrained Viterbi decoder, and estimates a confidence value for each complete gene model representing the probability that its entire structure is correct.
The system was trained across 2,055 RefSeq assemblies covering 2,047 eukaryotic species, including mammals, other vertebrates, invertebrates, plants, fungi, and protozoa. It is designed to annotate eukaryotic contigs and scaffolds without requiring external transcript evidence.
How it works
Carbon-A tokenises genomic DNA into non-overlapping 6-mers across an input context length of 16,384 tokens (98,304 base pairs). Longer sequences are processed in windows overlapping by 1,024 tokens (6,144 base pairs), with the overlapping probability predictions averaged. Two dedicated classification heads output per-base probabilities for non-coding and coding sequence on each strand independently.
The annotation pipeline processes these probabilities through structural decoding steps. A neural network scores potential gene and intron start and end boundaries, followed by a Viterbi decoder that identifies the optimal gene models conforming to biological grammar: canonical start codons, in-frame stop codons, valid intron splice pairs (GT-AG, GC-AG, or AT-AC), and frame maintenance across introns. A second network evaluates the completed gene model to assign a structural confidence score.
Key facts
- Parameter count
- 1.2 billion
- Architecture
- Transformer with two strand-specific classification heads
- Input format
- Genomic DNA in FASTA or GenBank (plain or gzip-compressed)
- Tokenisation
- Non-overlapping 6-mers (6 bp per token)
- Context window
- 16,384 tokens (98,304 base pairs)
- Primary outputs
- Gene annotations (GFF), CDS and protein sequences (FASTA), and per-base probabilities (Parquet)
- Training data
- 2,055 RefSeq GCF assemblies across 2,047 eukaryotic species
What you can use it for
- Eukaryotic genome annotation
- Annotate protein-coding genes in newly assembled or uncharacterised eukaryotic contigs and scaffolds.
- Strand-separated coding sequence detection
- Identify and resolve overlapping protein-coding genes located on opposite strands of DNA.
- Gene model confidence filtering
- Filter and prioritise predicted gene structures using predicted probabilities of structural correctness.
- Splice junction and frame resolution
- Derive structurally valid open reading frames across canonical GT-AG, GC-AG, and AT-AC introns directly from sequence.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
Carbon-A
Strengths and limits
Good at
- Generates independent forward- and reverse-strand probabilities, preventing the omission of genes overlapping on opposite strands.
- Provides a calibrated confidence score estimating the exact structural correctness of each predicted gene model.
- Enforces strict biological constraints via Viterbi decoding, including start codons, in-frame stops, and frame continuity across introns.
Watch out for
- Predicts protein-coding sequences only, omitting untranslated regions (UTRs) and non-coding RNAs.
- Outputs only one representative transcript isoform per gene rather than alternative splicing variants.
- Trained exclusively on eukaryotic nuclear genomes, making it unsuitable for bacteria, archaea, organelles, or viruses.
Inputs and outputs
You provide
Genomic DNA
FASTA, or a bare sequence. Up to a megabase per run, over at most 200 records. Give the model as much flank as you have: it reads a gene in the context of its neighbours.
Genetic codeoptional
The two tables Carbon-A's authors ship. Under the wrong one, genes break at codons the genome reads through, and the result says so.
You get back
One row per gene: record, strand, span, exon count, CDS and protein length, confidence, whether it is complete, its CDS coordinates and its translation.
Both strands' per-base coding probability, with the decoded exons drawn under each.
What was found, how to read the confidence, and GFF3 for a small result.
Licence
Commercial use permitted · Apache-2.0
Apache-2.0 on the model card for the weights, the modelling code and the annotation pipeline, and the repository is not gated. Its provenance.json names the source these weights were re-released from, GenerTeam/GENERanno-eukaryote-1.2b-cds-annotator-preview, which is MIT on the Hub and whose GitHub repository carries an MIT LICENSE file. Both licences permit commercial use. Read on 2026-10-29.
Questions
Can I use Carbon-A commercially?
Commercial use permitted. Apache-2.0 on the model card for the weights, the modelling code and the annotation pipeline, and the repository is not gated. Its provenance.json names the source these weights were re-released from, GenerTeam/GENERanno-eukaryote-1.2b-cds-annotator-preview, which is MIT on the Hub and whose GitHub repository carries an MIT LICENSE file. Both licences permit commercial use. Read on 2026-10-29.
How do I run Carbon-A without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 30s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
What inputs does Carbon-A require?
Carbon-A takes genomic DNA sequences in FASTA or GenBank format (which may be gzip-compressed) and an optional codon translation table setting.
Does Carbon-A identify non-coding genes or untranslated regions?
No. The model predicts only protein-coding sequence (CDS) and does not output UTRs or non-coding RNA features.
Can Carbon-A predict multiple transcript isoforms per locus?
No. The decoding pipeline generates a single gene model and isoform per predicted locus.
How are sequences longer than 98,304 base pairs processed?
Longer inputs are partitioned into 98,304-base-pair windows that overlap by 6,144 base pairs, with overlapping predictions averaged across windows.
Can the model annotate bacterial or viral genomes?
No. Carbon-A was trained solely on eukaryotic assemblies and is not designed for prokaryotes, viruses, or organellar DNA.
Written from Carbon-A's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.