Carbon-X

Carbon-X

Reads raw DNA inside a chat and answers in English, judging whether variants look benign or pathogenic, for researchers who want a DNA language model on their own GPU.

Made by HuggingFaceBio on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Carbon-X is a 14.8B language model from the Carbon family, trained on DNA sequences and English text together. You paste raw DNA into an ordinary chat and it replies in English, with its reasoning. It starts from Qwen3-14B as its base.

Who it suits: It suits researchers in genetics and biology who work with DNA sequences and want a model they can run on their own hardware.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Genetics research | Variant analysts | Judge whether a changed sequence looks harmless or disease-linked | Paste a reference and a variant, then ask for a call
Bioinformatics | Computational biologists | Sort sequences into organism groups | Ask which group a sequence belongs to
Genomics labs | Teams with unpublished data | Keep sequences on their own servers | Run the model on one GPU in the lab

If you do not work with DNA sequences or have no GPU, you can skip it.

Where to use it:

• Python scripts that use the transformers library
• A GPU machine, since the sample code loads the model on CUDA
• vLLM on a single GPU, which the card names for serving
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Variant effect | Geneticists | A benign or pathogenic call with an explanation
Gene control regions | Biologists | Regions flagged as steering gene activity
Organism group | Bioinformaticians | A group label for a sequence
Next bases | Researchers | A predicted continuation of a sequence

How to set it up: You need Python, a GPU with room for the weights, and Transformers version 5 or later.

• Python with pip
• A CUDA GPU, since the sample code uses it
• About 29.6 GB of disk space for the BF16 weights

1. Install or upgrade transformers.
2. Optionally save the files first with the two download commands.
3. Run the sample script. It loads the tokenizer with trust_remote_code turned on, because the DNA tokenizer is custom code.

pip install -U transformers

pip install -U huggingface_hub
hf download HuggingFaceBio/Carbon-X

How to use it: Put your DNA into a chat message and ask a question about it. The sample builds a message with a reference sequence and a variant, applies the chat template with thinking turned off, and prints the reply.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "HuggingFaceBio/Carbon-X"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="cuda")

messages = [
    {
        "role": "user",
        "content": (
            "Reference: ATGGTGCATCTGACTCCTGAGGAGAAGTCTGCCGTTACTGCCCTGTGGGGCAAGG\n"
            "Variant:   ATGGTGCATCTGACTCCTGTGGAGAAGTCTGCCGTTACTGCCCTGTGGGGCAAGG\n\n"
            "Is this variant likely to be benign or pathogenic? Explain your reasoning."
        ),
    }
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Title: Main calls in the sample
Tool | What it does
AutoTokenizer | Loads the hybrid text and DNA tokenizer
AutoModelForCausalLM | Loads the model onto the GPU in bfloat16
apply_chat_template | Turns your message into the model's prompt, with thinking off
generate | Writes the reply, up to the token limit you set

Example: You paste a reference sequence and a copy with one changed letter, then ask whether the change is likely benign or pathogenic and why. The model replies in English with a call and its reasoning, and the reply can quote DNA as well.

For newcomers and experts: Think of it as a chat assistant that can read DNA as well as English. A first small test is the sample script with one short sequence pair. Read the answer as a research aid, and check it against your own data.

DNA and text share one token stream in one decoder, with no separate DNA encoder. The hybrid tokenizer keeps the Qwen3 text vocabulary and adds 5,515 DNA tokens. Of those, 4,096 are 6-mers, so each DNA token stands for 6 base pairs. The model has 40 layers, hidden size 5120 and grouped-query attention with 8 key-value heads. It first saw 302B tokens of mixed DNA, biology and English data, then was tuned with supervised examples and reinforcement learning on DNA tasks.

Raw DNA inside a chat message
Hybrid tokenizer tags and splits it
One decoder reads DNA and text
English answer with reasoning

Good to know: The authors report an average of 66.7% across five DNA benchmarks with thinking off, 27 points above Qwen3-14B. They estimate that scoring 2.4 million ClinVar variants of uncertain significance would cost about $300, against $32,000 to $4 million for the frontier models they tested. The context window is 40,960 tokens. The card also has a section on using the model as an agent tool.

Title: Average score on five DNA benchmarks
Carbon-X: 66.7%
GPT-6.1 Sol: 67.5%

• License: Apache 2.0.
• Cost to run: none named, apart from a GPU. Per the card, one 80 GB GPU holds the bfloat16 weights.
• Data it reaches: the weights come from Hugging Face, and the card says sensitive or unpublished sequences can stay on your own infrastructure.
• Not for: extended reasoning. It is a chat model trained and tested with thinking off, and it needs Transformers 5 or later.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 14.8B
• Weights: about 29.6 GB in BF16; about 7.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,018
• Likes: 303
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3-14B
• Quantized versions: 14
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3, text-generation, genomics, biology, variant-effect-prediction, conversational

Links

• Model: https://huggingface.co/HuggingFaceBio/Carbon-X
• Files: https://huggingface.co/HuggingFaceBio/Carbon-X/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download HuggingFaceBio/Carbon-X

Then follow the usage in its model card. Nothing in this kit runs by itself.