Hugging Face's Carbon-X is an open 14.8B chat model that reads DNA

Hugging Face published Carbon-X, an open chat model that reads raw DNA and answers in English, on the Hugging Face Hub on October 22, 2026. The model card describes "a 14.8B-parameter decoder-only language model trained jointly on DNA sequences and natural-language text," released under Apache 2.0. Hugging Face presents it as a research model, and the card says: "Carbon-X is a starting point for research on connecting DNA and language. It has not been clinically validated, and it should not be used to make medical decisions."

#What changed?

Per the model card, "You give it raw DNA inside an ordinary chat conversation and it answers in English." Hugging Face lists these research capabilities: classifying DNA variants as benign or pathogenic on research benchmarks, identifying regions that control gene activity, classifying sequences by organism group, predicting the next bases of a sequence, and explaining its reasoning. "Responses can contain both DNA and natural language."

The key specs, from the card:

• Size: 14.8 billion parameters (WIRED and the release blog round it to 14 billion).
• Base model: "Carbon-X starts from Qwen3-14B."
• Design: "DNA and text share a single token stream in a single decoder. There is no separate DNA encoder." A hybrid tokenizer adds 5,515 DNA tokens to Qwen3's text vocabulary, with each DNA token covering 6 base pairs.
• Input: the card says it was trained on 8,192-token sequences, or about 49,000 base pairs of DNA, and accepts up to 40,960 tokens, but "we have not evaluated inputs much longer than those seen in training." WIRED reports it cannot take full human genomes.
• Training: about 300 billion tokens (302 billion per the model card) of DNA, biology text, English and instruction data, then supervised fine-tuning and reinforcement learning on DNA tasks. The mix is 55% raw DNA, 20% English, 15% biology text and 10% instruction data, according to the model card.
• Mode: "Chat model: post-trained and evaluated with thinking disabled."

The release blog says Carbon-X "can also act as a specialist sub-agent, helping other models interpret sequences they struggle to understand on their own." Hugging Face says that in its own agent evaluation, using Carbon-X as a sub-agent raised Kimi K3's DNA intelligence index from 49.4% to 60.5%. The card ships a ready-made skill for Claude Code and Codex CLI.

The card names the builders as "the research teams at Hugging Face, Zhongguancun Academy, and the Telethon Institute of Genetics and Medicine/University of Naples “Federico II”."

On safety, the blog says "we excluded eukaryote-infecting viral genomes from our DNA training data," and adds that "further research is needed to fully assess potential misuse risks, including after fine-tuning or integration with other systems."

#Who can use it and when?

Anyone. The repository is public and not gated, and the weights were published on October 22, 2026. Hugging Face lists these ways to run it:

Price: Hugging Face has not published one. The weights are free under Apache 2.0, and the Hub lists no hosted inference provider for Carbon-X.

The card says the full weights fit one GPU "so researchers can keep sensitive or unpublished sequences on their own infrastructure." The blog shows the model "running locally on a MacBook" with llama.cpp. Lewis Tunstall, a senior research scientist at Hugging Face, told WIRED that when it runs locally, no data is sent to Hugging Face or any other company. The card notes that hosted providers or Spaces may handle data differently.

#How to get it

1. Download the weights from the model repository. No sign-up gate applies.
2. For the full model, use a GPU with about 80 GB of memory and vLLM, with Transformers v5 or later. The card's serving command is vllm serve HuggingFaceBio/Carbon-X --trust-remote-code.
3. For smaller machines, use the GGUF builds with llama.cpp b11742 or newer, the first release with support for the Carbon DNA tokenizer, per the GGUF card. "Older builds load them but tokenize DNA wrongly." llama.cpp merged that tokenizer support on October 19, 2026.

#How does it compare?

All figures here are Hugging Face's own. The card says "Across five DNA benchmarks, which cover variant-effect prediction, sequence understanding and next-base prediction, Carbon-X averages 66.7% with thinking disabled," which it says is 27 points above Qwen3-14B. It puts GPT-6.1 Sol at 67.5% and says Carbon-X gets close "at about 1/1,800 of the inference cost per problem," using standard API prices and, for Carbon-X, an assumed $5 per H200-hour.

Hugging Face reports Carbon-X ahead on coding-variant effects (74.9) and next-base prediction (39.4). By its numbers, Carbon-X (85.1) trails GPT-6.1 Sol (91.7), Opus 5.5 (94.8) and Gemini 4 Argon (90.3) on complex variant effects, trails GPT-6.1 Sol and Opus 5.5 on single-base variant effects, and trails GPT-6.1 Sol on GenomeQA. The card says "BioReason and GenomeQA come from the literature; we built FrontierVEP and NextBase," so Hugging Face built two of the five benchmarks. We found no independent evaluation. Hugging Face also reports 79.3% on IFEval, prompt-level strict.

Hugging Face estimates that running all 2.4 million variants of uncertain significance in the ClinVar database through Carbon-X would cost about $300, against $32,000 to $4 million with the frontier models it tested.

For other context, WIRED notes that Google DeepMind's AlphaGenome gives its predictions "in a numeric format" and that Evo 2 from the Arc Institute and Nvidia "can generate entire genomes from scratch." Carbon-X differs: the card says "Carbon-X reads and explains existing sequences. It was not trained to design new ones."

DeepSeek V4.1 Flash is among the models Hugging Face compares against; see our DeepSeek V4.1 Flash coverage and our OpenAI DevDay 2026 recap for GPT-6.1 Sol. The blog says "As frontier labs move into biology, researchers should not have to depend on a handful of closed APIs"; for one such move, see OpenAI's acquisition of Orchestra Bio. For another model that runs on your own machine, see Perplexity Computer running fully local on Macs.

#What is still unclear?

• Independent results. Every score is Hugging Face's, and no paper or technical report for Carbon-X was found.
• Lab validation. WIRED reports the developers "haven’t validated the model yet with actual wet lab experiments, so they can’t say how accurate its predictions are."
• Known errors. The card says "The model can miscount bases, confuse coordinates or make incorrect edits, particularly when it locates exact positions in long sequences," and that it "may invent a sequence's identity or function when the prompt does not give enough information."
• Misuse risk. Hugging Face says more research is needed, including after fine-tuning.