DNA-language chat model that reads raw genomic sequence in a conversation and answers in English, scoring variant effects and regulatory elements.
No providers recorded yet. Browse all providers
A genomic language model can score a mutation but cannot say why, and a general-purpose chat model can explain what a gene does but has seen almost no raw sequence. Carbon-X is built to close that gap. It is a single 14.8B-parameter decoder that reads DNA and English in one token stream, so a researcher can paste a reference and variant sequence into a chat and get a classification with a prose explanation.
Carbon-X was released in October 2026 by Hugging Face with the Zhongguancun Academy and the Telethon Institute of Genetics and Medicine / University of Naples Federico II, the same collaboration behind the Carbon genomic foundation models. Unlike the Carbon models, trained from scratch on DNA, Carbon-X starts from the general-purpose Qwen3-14B and learns DNA as a new modality. Earlier DNA–language systems such as BioReason pair a separate DNA encoder with a language-model decoder; Carbon-X has no separate encoder, and DNA tokens are simply additional vocabulary in the same model. The tokenizer design follows Carbon's 6-mer results and the approach of NatureLM to mixing biological sequence with text.
Carbon-X is documented through its model card, a release blog post and an open training repository. There is no preprint or peer-reviewed paper, and all benchmark results are the developers' own.
<dna>, </dna> and <oov> markers.Carbon-X keeps the Qwen3-14B architecture: 40 layers, hidden size 5,120, grouped-query attention with 40 query and 8 key-value heads, SwiGLU, RMSNorm with QK-norm, RoPE and a 40,960-token context window. The new DNA embeddings were first trained for about 10B tokens with the transformer body frozen. Mid-training then ran for 302B tokens at a sequence length of 8,192: 55% raw DNA (eukaryotic genomes from the GENERator corpus plus OpenGenome2 mRNA and prokaryotic genomes), 20% English web and PDF text, 15% biology text including documents containing long DNA sequences, and 10% instruction data. Supervised fine-tuning followed on a 20-subset mixture led by variant-effect reasoning, and then two rounds of asynchronous GRPO with task-specific verifiers. The mid-trained (Carbon-X-Base) and fine-tuned (Carbon-X-SFT) checkpoints are released alongside the final model, and the full training data is published as the FineDNA dataset.
On five DNA benchmarks, Carbon-X averages 66.7% with thinking disabled, 27 points above Qwen3-14B. It scores 74.9 balanced accuracy on BioReason Coding (coding variants with gene names removed), 85.1 on BioReason Non-SNV (indels and other complex variants), 63.6% on GenomeQA (multiple-choice sequence-feature questions), 39.4% per-base accuracy on NextBase and 70.4% on FrontierVEP, which pairs a single-nucleotide variant with up to 8,192 bases of reference context. FrontierVEP and NextBase were built by the Carbon-X team. Training data was decontaminated against all five benchmarks by DNA overlap, text n-gram overlap and homology search. The model reaches 79.3% prompt-level strict accuracy on IFEval.
Carbon-X is aimed at researchers who want to interrogate sequences conversationally: triaging variants of uncertain significance, distinguishing enhancers from promoters, identifying transcription-factor motifs and binding sites, classifying sequences by organism group, and continuing a sequence by a few bases. Its low per-query cost suits batch jobs; the developers estimate about $300 to run all 2.4 million ClinVar variants of uncertain significance. The release also ships tool wrappers and a coding-agent skill so a general-purpose assistant can hand it DNA questions; requiring Kimi K3 to consult Carbon-X raised its five-benchmark average from 49.4% to 60.5%.
Carbon-X shows that an open general-purpose language model can absorb raw genomic sequence as a second modality while keeping its English, and the finding that reinforcement learning on Qwen3-14B alone stayed at chance on FrontierVEP until DNA mid-training came first is useful to anyone attempting the same adaptation. The release is unusually complete: weights for all three stages, training data, recipe and evaluation code are public under Apache 2.0. The stated limits are specific. The model is not clinically validated, remains weak on splice-site (29.8%) and histone-mark (34.6%) questions, can miscount bases or confuse coordinates in long sequences, and was not trained to design new DNA. Inputs far beyond the 8,192-token training length are unevaluated.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.