Hugging Face open-sources Carbon-X, a 14B Apache-2.0 model you can chat with about DNA: averages 66.7% on five DNA benchmarks vs GPT-6.1 Sol's 67.5% at ~1/1,800 the cost, runs on one 80 GB GPU

Why it matters

【Key points】

On Oct 22, Hugging Face's biology team (HuggingFaceBio), with Zhongguancun Academy and the University of Naples, released Carbon-X under Apache 2.0. It is a 14.8B decoder built from Qwen3-14B with 5,515 added DNA tokens (6-mers). DNA and text share one token stream. It was mid-trained on 302B tokens (55% DNA, no viral genomes), then SFT and RL with verifiable rewards.

Paste reference and variant sequences into a chat and it says whether a mutation is benign or pathogenic, and explains why. On five DNA benchmarks with thinking off it averages 66.7%, against GPT-6.1 Sol 67.5%, Opus 5.5 60.9%, Kimi K3 50.4% and Gemini 4 Argon 49.8%. It is +27 points over Qwen3-14B, at about 1/1,800 of Sol's cost per problem. The team says screening ClinVar's 2.4M variants of uncertain significance would cost ~$300 vs $32K–$4M on frontier models. It fits one 80 GB GPU.

【Verifiable】

https://huggingface.co/HuggingFaceBio/Carbon-X

https://huggingface.co/spaces/HuggingFaceBio/Carbon-X-Release

https://www.wired.com/story/hugging-face-new-ai-biology-model-chat-with-dna/

https://www.businessinsider.com/hugging-face-ai-model-carbon-x-dna-sequences-gene-therapies-2026-10

https://agipulse.org/e/817

【Our read】

1. It joins sequence-only genomic LMs (Evo 2, Nucleotide Transformer) and numeric predictors (AlphaGenome) with a chat model. It is the first open "talk to DNA" model near frontier scores; Carbon-A came Oct 8.

2. The near-tie with Sol is real but narrow. Two of the five benchmarks are self-built, frontier models still lead on non-SNV variants (94.8/91.7 vs 85.1), and splice-site and histone tasks sit near 30%.

3. The shift is cost and privacy. Labs can screen sensitive variants on one GPU, and specialist models can serve as tools: Kimi K3 rose from 49.4% to 60.5% when it called Carbon-X.

4. Caveats: it is not clinically validated, reliable input is ~50 kbp, it is weak at long DNA generation despite some headlines, and all scores are self-reported.