Blog | Hugging Face Bio | Code

Carbon-X GGUF

GGUF quantizations of Carbon-X, an open 14B language model for DNA understanding and generation from the Carbon family, for local inference with llama.cpp.

Carbon-X reads raw DNA inside an ordinary chat conversation and answers in English: it predicts whether mutations are benign or pathogenic, identifies regions that control gene activity, classifies sequences by organism group, predicts the next bases of a sequence and explains its reasoning. A hybrid tokenizer keeps a BPE vocabulary for text and encodes DNA as non-overlapping 6-mers. See the Carbon-X model card for details on the model and its limitations.

[!IMPORTANT] These files need llama.cpp b11742 or newer, the first build with support for the Carbon HybridDNATokenizer (#30825). Older builds, including the v0.7.0 release, load them but tokenize DNA wrongly.

Table of Contents

1. Available files
2. How to use
3. Evaluation
4. License
5. Acknowledgements

Available files

Q8_0 is the recommended file: half the size of BF16, with the same benchmark scores. Use Q4_K_M when memory is tight. In every K-quant the token embedding and output tensors, which hold the DNA tokens, stay at Q8_0.

How to use

Click Use this model > llama.cpp on this page for a ready-made snippet. The easiest way to run it is the llama.cpp desktop app for macOS and Windows, which lists every model in your Hugging Face cache. Alternatively, install llama.cpp from llama.app or download a prebuilt binary from the releases page, then run:

llama serve -hf HuggingFaceBio/Carbon-X-GGUF:Q8_0   # opens a chat web UI and an OpenAI-compatible API
llama cli -hf HuggingFaceBio/Carbon-X-GGUF:Q8_0     # chat in the terminal

No other flags are needed: the GGUF files carry the chat template and Carbon-X's default sampling settings. Carbon-X was post-trained and evaluated with thinking disabled. To match that, add --chat-template-kwargs '{"enable_thinking": false}'.

Working with DNA inputs

The tokenizer switches between BPE for text and 6-mers for DNA. Paste sequences directly into a message:

• DNA is tagged automatically. Runs of at least 6 A/C/G/T bases (either case) are wrapped in <dna>…</dna> and tokenized as 6-mers. Runs of at least 12 bases with up to 20% IUPAC ambiguity codes (N, R, Y, …) are tagged too, and sequences displayed in space-separated blocks are joined. Text you have already wrapped in <dna>…</dna> is always tokenized as DNA.
• Only A, C, G and T form 6-mers. A 6-mer that contains any other character, such as N, becomes a single <oov> token. To keep full resolution, filter to canonical bases first.
• Any length works. When a sequence's length is not a multiple of 6, its last 1–5 bases are encoded with dedicated tail tokens, and decoding returns the original sequence exactly.
• DNA in responses is tagged. The model writes DNA inside <dna>…</dna>, and llama.cpp keeps the tags in the returned text. Strip them if you need the bare sequence.

The llama.cpp tokenizer matches the Python HybridDNATokenizer exactly: identical token ids, detokenized text and chat-template rendering on 1,000 benchmark prompts (1.13M tokens) and 36 DNA edge cases.

Prompting tips

• Put the sequence and the question in the same user message.
• Include any context you already have, such as the gene, strand, coordinates or annotations. Without it, the model may guess at a sequence's identity or function.
• For variant questions, give the reference and alternate sequences, or a reference sequence with the position and the base change.

Evaluation

We ran two of the DNA benchmarks from the Carbon-X model card on each GGUF file with the llama.cpp server, with thinking disabled, Carbon-X's default sampling settings and one sample per problem. The Carbon-X column is the original model served with vLLM.

• FrontierVEP: given a long reference sequence and a single-nucleotide variant, predict whether the variant is benign or pathogenic (500 benign, 500 pathogenic).
• NextBase: predict the next DNA bases of a sequence. The score is the fraction of bases predicted correctly.

Scores are single runs with sampling: the standard error is about 1.5 points on FrontierVEP and 1 point on NextBase. BF16 through Q5_K_M are within about one standard error of Carbon-X. Q4_K_M is the lowest on both benchmarks.

License

Apache 2.0.

Acknowledgements

Carbon-X is a joint collaboration between the research teams at Hugging Face, Zhongguancun Academy, and the Telethon Institute of Genetics and Medicine/University of Naples “Federico II”.