Hugging Face's Carbon-X: an open 14B model that reads raw DNA and answers in English, close to GPT-6.1 Sol on its own DNA benchmarks

Hugging Face has released Carbon-X, a 14.8B-parameter language model that takes raw DNA inside an ordinary chat and answers in English. It predicts whether a mutation is benign or pathogenic, finds regulatory regions, classifies sequences by organism group and predicts the next bases of a sequence, explaining its reasoning as it goes. It starts from Qwen3-14B and is released under Apache 2.0 together with the full training data, FineDNA, and the training recipe. It is a joint project of Hugging Face, Zhongguancun Academy and the Telethon Institute of Genetics and Medicine / University of Naples "Federico II". The FineDNA citation lists Loubna Ben Allal, Lewis Tunstall, Dana Aubakirova, Maurizio Fiusco, Edward Beeching, Kashif Rasul, Leandro von Werra, Qiuyi Li and Diego di Bernardo.

The headline claim from the model card: across five DNA benchmarks Carbon-X averages 66.7% with thinking disabled. That is 27 points above Qwen3-14B and close to GPT-6.1 Sol at 67.5%, at what Hugging Face estimates is about 1/1,800 of the inference cost per problem. Claude Opus 5.5 scores 60.9% on the same set. Carbon-X leads every model tested on coding-variant effects (74.9 on BioReason Coding) and next-base prediction (39.4 on NextBase).

How it's built

There is no separate DNA encoder. DNA and text share one token stream in one decoder. The tokenizer keeps Qwen3's BPE vocabulary for text and adds 5,515 DNA tokens: all 4,096 non-overlapping 6-mers, 1,364 tail tokens for sequences whose length isn't a multiple of 6, and a few markers. DNA in a prompt is detected and tagged automatically, so one DNA token covers six base pairs. The context window is 40,960 tokens.

Training ran in three stages, and all of the data is public in FineDNA:

• Mid-training with Megatron-LM on 302B tokens (read out of a 514B-token pool built from 23 sources): roughly 55% raw DNA, 20% English, 15% biology text and 10% instruction data, starting with a phase where only the new embeddings train.
• SFT with TRL on 203,923 conversations (230M tokens) across eight domains.
• RL with asynchronous GRPO and verifiable rewards on DNA tasks, using 98,433 prompts.

The weights fit on one 80 GB GPU in bfloat16 and serve with vLLM. There are also Carbon-X-Base, Carbon-X-SFT and a GGUF build, plus an agent skill for using it from Claude Code and Codex. The recipe is in huggingface/carbon.

The cost argument

The most concrete number on the card isn't a benchmark score. Hugging Face estimates that running Carbon-X over the 2.4 million variants of uncertain significance in ClinVar would cost about $300, against $32,000 to $4 million with the frontier models it tested. Its cost figure assumes $5 per H200-hour at measured throughput, while the frontier models use standard API prices. Because the model runs locally, labs can also keep unpublished or patient sequences on their own hardware, which the card points out explicitly.

What to keep in mind

• Who it's compared with. The comparison table is against general LLMs (GPT-6.1 Sol, Opus 5.5, open chat models), not against specialist DNA models such as Evo 2 or Nucleotide Transformer. Ben Allal says that's deliberate: specialist models like PlantCAD2 "give you a score", while Carbon-X is meant to reason about a sequence in plain English first, and "specialized models are still great for scoring whole genomes cheaply."

• Whose benchmarks. Some of the five benchmarks, including NextBase, where Carbon-X leads, were built by the same team for this release.
• Weak spots. By the card's own numbers, splice-site and histone-mark questions are weak (29.8% and 34.6% on GenomeQA). The model can miscount bases or confuse coordinates in long sequences, may invent a sequence's function when context is missing, and degrades when asked to generate long DNA.
• Not clinical. Hugging Face says plainly that Carbon-X "has not been clinically validated, and it should not be used to make medical decisions."

The open question is how Carbon-X does against the specialist DNA models on shared, outside benchmarks. With weights, data and recipe all public, that's something anyone can now check.