🧬 Open Beta

DNA Model Benchmarking

Compare HuggingFaceBio Carbon foundation models on your sequence. Select multiple models, run inference, and export results.

~30s first call (model cold start)

FASTA upload supported

No account required

Configure run

DNA sequence required

Only A, T, G, C, U characters. Lines and whitespace are stripped automatically.

Or upload FASTA file

Carbon models ?

Carbon-500M 0.5B params

Fast draft model. Use with Carbon-3B or Carbon-8B as target for speculative decoding — same quality, lower compute cost. Shared 6-mer tokenizer.

Carbon-3B 3B params

Flagship model. Matches Evo2-7B performance at 150× faster throughput. 6-mer tokenizer, YaRN-extendable to 393 kbp context. Best value/speed trade-off.

Carbon-8B 8B params

Largest Carbon checkpoint. Highest quality across all benchmarks. Same 6-mer tokenizer and YaRN context extension as Carbon-3B.

Max new tokens 6-mers

8 64 256

1 token ≈ 6 DNA bases. At 64 tokens, max output is ~384 bases.

Temperature

0.1 (deterministic) 0.7 2.0 (creative)

First call triggers model loading (~30s). Subsequent calls are faster. Results are streamed from HuggingFace Inference API.

Benchmark scores (Carbon-3B)

Task	Score
Sequence Recovery (eukaryote)	61.54%
BRCA2 VEP	84.63%
ClinVar Coding (24kb)	92.89%
ClinVar Non-coding (24kb)	91.14%
Nucleotide triplet-expansion	85.20%
Synonymous codon replacement	88.89%

Zero-shot scores from HuggingFaceBio. View full report →

Ready to benchmark

Select a Carbon model, paste a DNA sequence, and click Run Benchmark.

Models load fresh from HuggingFace Hub on each request.