Hugging Face’s biology team has published Carbon-X, an open-weight model that accepts raw DNA sequences inside a chat prompt and replies in plain English. According to its model card, the 14.8-billion-parameter model is released under the Apache 2.0 license and scores an average of 66.7% on five DNA benchmarks, 0.8 points below the 67.5% the card lists for OpenAI’s GPT-6.1 Sol. All figures below come from Hugging Face’s own documentation and have not been independently verified.
What Carbon-X is
Carbon-X is a decoder-only chat model that belongs to the Carbon family from HuggingFaceBio, the company’s biology research group. The model card says it starts from Qwen3-14B and is trained on both DNA and natural-language text, so a single model can read a sequence and discuss it in words. The Hub lists the size as 15B, while the card text gives 14.8B. Its context window is 40,960 tokens, and weights are in bfloat16.
The design choice that sets it apart is a hybrid tokenizer. Text uses ordinary byte-pair encoding, while DNA is split into non-overlapping 6-mers, so one token covers six base pairs. DNA is detected and wrapped in <dna> tags automatically, and only the letters A, C, G and T are turned into 6-mers. Using it requires trust_remote_code=True.
How it was trained
The card describes four stages: adding DNA tokens to the vocabulary, mid-training on 302 billion tokens (55% raw DNA, 20% English text, 15% biology text, 10% instructions), supervised fine-tuning, and reinforcement learning with verifiable rewards. It also says training data was checked against the benchmarks for overlap.
What the benchmarks show
Hugging Face compares Carbon-X with GPT-6.1 Sol, Opus 5.5, Gemini 4 Argon, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, all with thinking disabled and one sample per problem. The frontier models were run at the effort levels the card names, such as “Max” or “Low.”
| Benchmark | Carbon-X | GPT-6.1 Sol | Opus 5.5 |
|---|---|---|---|
| BioReason Coding (variant effects) | 74.9 | 71.0 | 70.3 |
| BioReason Non-SNV | 85.1 | 91.7 | 94.8 |
| GenomeQA | 63.6 | 67.2 | 62.2 |
| NextBase (next-base prediction) | 39.4 | 32.9 | 1.1 |
| FrontierVEP | 70.4 | 74.9 | 76.2 |
| Average | 66.7 | 67.5 | 60.9 |
Carbon-X leads on coding-region variants and next-base prediction but trails the two top closed models on complex and single-base variant effects. The authors built FrontierVEP and NextBase themselves; BioReason and GenomeQA come from earlier literature.
The cost claim
The card says Carbon-X is near GPT-6.1 Sol’s average at roughly 1/1,800 of its per-problem inference cost. It estimates about $300 to analyze the 2.4 million variants of uncertain significance in ClinVar, against $32,000 to $4 million for the frontier models it tested. These estimates assume an H200 GPU at $5 per hour. The model runs on a single 80 GB GPU through vLLM, SGLang, Docker Model Runner or Transformers v5 or later; GGUF runtimes such as llama.cpp are not supported because of the custom tokenizer.
What it cannot do
The card is explicit about limits:
- It has not been clinically validated and should not guide medical decisions.
- It is unreliable on exact operations such as counting, coordinates and edits, and can hallucinate a sequence’s identity or function when context is missing.
- It is weak on some GenomeQA tasks, with 29.8% on splice sites and 34.6% on histone marks.
- Quality drops on long DNA generation, and it was not trained to design sequences.
- Genetic data is sensitive, and its handling depends on whoever hosts the model.
For wider context, see our look at how AI is accelerating science and the Google Research wet-lab binder study.
What to watch next
- Independent evaluations of Carbon-X on benchmarks that Hugging Face did not build itself.
- Whether outside labs reproduce the card’s results using the evaluation code in the Carbon repository.
- Any clinical or research groups that test it on real variant-interpretation workflows and report back.
Key takeaways
- Carbon-X is an Apache 2.0, open-weight 14.8B model built on Qwen3-14B that takes DNA and English in one prompt.
- Hugging Face reports a 66.7% five-benchmark average versus 67.5% for GPT-6.1 Sol; the numbers are self-reported.
- It wins on coding variant effects and next-base prediction but trails on complex variant effects.
- The model is not clinically validated and is weak at exact operations and sequence design.
- It fits on one 80 GB GPU, which puts genomic question answering within reach of small labs.
Frequently asked questions
Is Carbon-X really open?
The model card lists the Apache 2.0 license, and the weights are on the Hugging Face Hub. That fits the usual meaning of an open-weight model, though the card does not describe the full training data release.
What can Carbon-X do with DNA?
Per the card, it can predict whether variants are benign or pathogenic, tell enhancers from promoters, classify sequences by organism group, and predict next bases, explaining its reasoning in English.
Can it be used for medical decisions?
No. The card says it has not been clinically validated and should not guide medical decisions.
