Hugging Face's Carbon-X reads DNA like a language, answers in English

How it works

A free AI model you run on your own computer that reads DNA letters and explains them in plain English.

1. 💻Upload your DNA data onto your own GPU
2. 🗣️Ask a question about the genes in English
3. 📖Get a plain-English answer back instantly

What you could do with it

• A biotech researcher could chat with raw DNA sequences on their own GPUs and get plain-English answers about mutations, for a tiny fraction of the cost of API-based tools ¹.
• A clinical lab could run carbon-X on sensitive genetic data without sending it to a third party, keeping everything on their own GPUs under an open-source license ¹.
• A biology student could predict whether a specific mutation is benign or harmful, flag regulatory regions, or identify the next base in a DNA strand using a single open model ¹.

An exampleInvented to explain it

Priya, a grad student studying genetic mutations

PriyaIs this DNA change harmless or dangerous?

Carbon-XThat mutation appears in a regulatory region and may be harmful.

PriyaWhat part of the gene is affected?

Carbon-XIt sits near a splice site that controls how the gene is read.

PriyaCan I trust this result?

Carbon-XMy confidence is moderate; experimental validation is recommended.

She saved her analysis locally and avoided paying API fees entirely.

A new kind of model

Carbon-X is built on top of Qwen3-14B, an open language model, but with a special twist. Instead of treating DNA and words as separate things, it streams them together using a hybrid tokenizer. The tokenizer keeps Qwen3's usual vocabulary for English and adds 5,515 DNA-specific tokens. Each DNA token represents six base pairs — the letters A, C, G and T that make up the genetic code ¹.

The model has 40 layers and a hidden size of 5,120. It uses a standard decoder-only architecture, meaning it does not rely on a separate DNA encoder like some earlier tools did ¹. One DNA token covers six base pairs, and any run of six or more A, C, G or T bases is tagged automatically .

How it learned to read DNA

Training happened in four stages. First came an embedding warmup using about 10 billion tokens with the main model frozen, letting only the new DNA pieces learn. Next was mid-training on 302 billion tokens at a sequence length of 8,192, mixing 55 percent raw DNA, 20 percent English text, 15 percent biology writing and 10 percent instruction data ¹.

The final stages used supervised fine-tuning over 20 datasets and reinforcement learning called GRPO, which gives the model rewards when its answers pass checks built by domain experts ¹. The hardest part, per the release post, was picking the right learning rate. High rates taught DNA but made the model forget English; a lower rate over a longer run fixed the trade-off .

What it can do

Carbon-X scores 66.7 percent on average across five DNA benchmarks, trailing GPT-6.1 Sol by just 0.8 points ¹. It leads every tested model on BioReason Coding at 74.9 percent and on NextBase prediction at 39.4 percent . Its weakest score is 29.8 percent on GenomeQA splice-site questions .

The model keeps general reasoning skills too, hitting 79.3 percent on the IFEval prompt-following benchmark ¹. It can predict whether a mutation is harmless or harmful, flag regulatory regions, classify organisms and guess the next DNA base .

Critically, the scores are self-reported by Hugging Face, and Carbon-X ran with thinking disabled while rivals ran at their highest reasoning effort ¹. On benchmarks like NextBase, some competitor models actually scored worse when they tried harder because refusals went up .

Guess

How much would scoring 2.4 million variants cost on frontier APIs?

$32,000 to $4 million$300About $5 total

Price and access

The model ships under the Apache 2.0 license, so labs can keep sensitive genetic sequences on their own hardware ¹. It runs on a single 80 GB GPU in bfloat16 through vLLM or the Transformers library, and an official GGUF version comes in at 15.8 GB (Q8_0) or 9.6 GB (Q4_K_M) .

Hugging Face prices Carbon-X at about $5 per hour on an H200 GPU, which works out to roughly 1/1,800th the cost of GPT-6.1 Sol per problem ¹. Scoring 2.4 million uncertain genetic variants from ClinVar would cost around $300, compared with $32,000 to $4 million on frontier APIs .

Kimi K3 also benefits from the model: its DNA index rose from 49.4 percent to 60.5 percent when it consulted Carbon-X ¹.

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

How good is it

According to Hugging Face, Carbon-X averages 66.7 percent across five DNA benchmarks, trailing GPT-6.1 Sol by just 0.8 points ¹. It leads every tested model on BioReason Coding at 74.9 percent and on NextBase prediction at 39.4 percent . Its weakest result is 29.8 percent on GenomeQA splice-site questions . These scores are self-reported by Hugging Face and were collected with thinking disabled, while competitors ran at their highest reasoning effort .

percentage of correct answers on DNA-specific tests. Higher is better. Numbers: Hugging Face (self-reported). Chart drawn by Herald of AI.

Key facts

• 14.8B parameters, 40,960-token context, runs on 1x 80GB GPU

It scored an average of 66.7% across 5 DNA benchmarks, placing it 0.8 points behind GPT-6.1 Sol

• Leads all tested models on BioReason Coding at 74.9%
• Costs ~1/1,800th of GPT-6.1 Sol per inference problem
• Apache 2.0 license; sensitive DNA stays on your own hardware

Sources

1. Hugging Face Releases Carbon-X, a 14B Open DNA Language Model that Brings DNA and Natural Language Together for the Firs — marktechpost.com