Hugging Face Carbon-X: what it is and how to run it
2026-10-25 · 5 min read
Carbon-X is a free, open-weight model from Hugging Face. You paste in raw DNA, ask a question in plain English, and it answers. It has 14.8 billion parameters and an Apache 2.0 license, so you can run it on your own hardware: in full precision on one 80 GB GPU, or as a 9.6 to 15.8 GB quantized file through llama.cpp.
Hugging Face researcher @_lewtun announced it on October 22, calling it "a 14B model that brings DNA and natural language together for the first time." His colleague @LoubnaBenAllal1 went with "the best open DNA model" and suggested three uses: ask about a sequence, ask what a mutation does, or have it generate new DNA.
What the X posts claim
The thread from @_lewtun makes five claims. All of them are Hugging Face's own numbers:
- It beats every open language model the team tested on DNA benchmarks, and lands within a point of OpenAI's GPT-6.1 Sol at what he puts at over 1,000 times lower inference cost.
- It was mid-trained on about 300 billion tokens of DNA mixed with text, then fine-tuned and trained with RL, without distilling from another model.
- A new hybrid tokenizer stores DNA in about a third of the tokens a standard tokenizer needs.
- Used as a DNA specialist subagent, it lifted Kimi K3's scores by 11.1 points.
- It runs locally with llama.cpp.
The model card has the numbers behind the first claim. Across five DNA benchmarks Carbon-X averages 66.7, against 67.5 for GPT-6.1 Sol and 60.9 for Claude Opus 5.5. The card estimates the cost per problem at about 1/1,800 of GPT-6.1 Sol's. Carbon-X wins outright on two tasks, protein-coding questions (74.9 vs 71.0) and next-base prediction (39.4 vs 32.9), and trails on the other three. TNW points out that nobody outside Hugging Face has verified any of these results yet.
The model is built on Alibaba's Qwen3-14B, with a 5,515-token DNA vocabulary added. TNW reports that Tunstall said the team left viruses out of the training data to reduce the risk of misuse.
How to run Carbon-X
There are three routes, and which one fits depends on your hardware.
| Route | What you need | Cost |
|---|---|---|
| Transformers or vLLM, full weights | One 80 GB GPU, Transformers v5+ | Free weights, your GPU bill |
| llama.cpp, Q8_0 GGUF | About 16 GB of memory, llama.cpp b11742 or newer | Free |
| llama.cpp, Q4_K_M GGUF | 9.6 GB file, for tight memory | Free |
On a GPU server
The model card gives a one-liner for an OpenAI-compatible endpoint:
pip install -U vllm
vllm serve HuggingFaceBio/Carbon-X --trust-remote-code
You need --trust-remote-code because the DNA tokenizer is custom code that ships with the model.
On a workstation or a Mac
The official Carbon-X-GGUF repo recommends the Q8_0 file and serves it like this:
llama-server -hf HuggingFaceBio/Carbon-X-GGUF:Q8_0 -ngl 99 -c 40960 --jinja --port 8080
Use llama.cpp build b11742 or later. The repo warns that older builds will load the file but tokenize DNA wrong, and you won't get an error to tell you. Keep thinking mode off. The GGUF page's full command passes enable_thinking: false and sets its own sampling values. In Hugging Face's own tests, everything from BF16 down to Q5_K_M scored about the same as the full model, and Q4_K_M came in lowest.
One inconsistency: the Agent section of the main model card still says GGUF runtimes can't load the tokenizer, while the GGUF repo and @_lewtun's post say they can. My guess is the card predates the llama.cpp build that added support. Either way, try it on a sequence you already know the answer to before you rely on it.
What to ask it
Put the sequence and the question in one message, and add the gene, strand or coordinates if you have them. Any run of six or more A/C/G/T letters gets tagged as DNA automatically. Two prompt shapes from the model card:
- "What does this sequence encode?" followed by the sequence.
- A reference sequence and a variant sequence, then "Is this variant benign or pathogenic?"
What it can't do
Carbon-X has not been clinically validated and should not be used for medical decisions. It can miscount bases and mix up coordinates, so use code for anything that has to be exact. Without context it may make up what a sequence is. It scores 29.8% on splice-site questions and 34.6% on histone marks, and it wasn't trained for long sequence design. Genetic data is also sensitive, so check a hosted demo's data terms before you paste anything real into it.
My read
I care most about the subagent claim. In that setup a frontier model runs the conversation and passes DNA questions to a small, cheap specialist on your own machine. Kimi K3 was the main model in Hugging Face's test (we covered the newer Kimi K3.1 Pro this week). The small-model half looks a lot like laptop-sized coders such as Recursive Flint. In fields with private data, that split matters, because the sensitive records stay on hardware you control.
Most businesses will never work with a genome. The same setup, a big general model plus a small specialist for one narrow job, applies to any business with private records, and it only pays off once the workflow around it is written down. If you want to know where a specialist could fit in yours, New Face Design's free process audit maps that out.