Hugging Face Carbon-X: An Open 14B Model You Can Chat With About DNA

Hugging Face has released Carbon-X, an open 14.8-billion-parameter model that lets researchers paste raw DNA into a chat and get an answer in plain English. The model card on the Hub lists it under Apache 2.0, and RuntimeWire, citing Wired, reported the launch on October 22, 2026, noting that wet-lab validation has not been done. The headline claim from the card: across five DNA benchmarks Carbon-X averages 66.7%, within a point of GPT-6.1 Sol's 67.5%, at about 1/1,800 of the inference cost per problem.

That is a big claim from the people who built the model, so this post separates what the model card documents, what is a company estimate, and what nobody has independently checked yet. It follows Carbon-A, the gene-finding model Hugging Face released on October 13, and is the second Carbon release in nine days.

TL;DR: Carbon-X in one table

What does Carbon-X actually do?

A rounded double helix with one green rung, illustrating how Carbon-X reads DNA sequences in chat

You give the model a reference sequence and a variant, or a stretch of DNA and a question, and it replies in language. The card lists the tasks it was trained for: predicting whether a mutation is benign or pathogenic, finding regions that control gene activity, classifying sequences by organism group, predicting the next bases of a sequence, and explaining its reasoning.

The design choice that makes this work is a single token stream. There is no separate DNA encoder bolted onto a language model. A hybrid tokenizer keeps the Qwen3 text vocabulary and adds 5,515 DNA tokens: 4,096 six-letter "6-mers", 1,364 tail tokens for sequences whose length is not a multiple of six, and the <dna>, </dna> and <oov> markers. One DNA token therefore covers six base pairs. Runs of at least six A, C, G or T bases in a prompt are detected and tagged automatically.

Anything outside A, C, G and T is a problem: a 6-mer that contains an N becomes a single out-of-vocabulary token, so the card advises filtering to canonical bases first. Sequences up to the context limit work, but inputs much longer than those seen in training have not been evaluated.

How was it trained?

The recipe is more interesting than the headline, because it shows how a general model was converted into a specialist without forgetting English.

1. Vocabulary extension. The new DNA embeddings were randomly initialized, then trained for about 10 billion tokens with the transformer body frozen so only the new embeddings learned.
2. Mid-training. 302 billion tokens at sequence length 8,192: 55% raw DNA (eukaryotic genomes, mRNA, prokaryotic genomes from OpenGenome2), 20% English web text, 15% biology text and 10% instruction data. The authors say learning rate was the dominant knob: too low leaves DNA at chance, too high learns DNA but forgets English.
3. Supervised fine-tuning. Three epochs on a mixture of 20 dataset subsets, about 30% of it variant-effect reasoning over several views of the same variant.
4. Reinforcement learning with verifiable rewards. Two rounds of asynchronous GRPO, using code-checked verifiers for variant-effect prediction, sequence classification and basic DNA operations, with a penalty for ungrounded rationales.

The card says all training data was decontaminated against the benchmarks using whole-span DNA overlap, text n-gram overlap and sequence homology search. If you want the background on why verifiable rewards are suddenly everywhere, our look at reinforcement learning making models stop speaking English covers the trend from the language side.

What the benchmarks say (company-reported)

The table below is Hugging Face's own, copied from the card. Carbon-X runs with thinking disabled and one sample per problem. Each rival is shown at the reasoning effort that gave it the highest average. BioReason and GenomeQA come from the literature; the authors built FrontierVEP and NextBase themselves.

The card also reports GLM-5.3 (39.8), Kimi K3 (50.4) and DeepSeek V4.1 Flash (47.1) averages, all at low effort. Read the table honestly: Carbon-X wins on coding-variant effects and next-base prediction, and frontier models still win on non-SNV variants (insertions, deletions) and on single-base variant effects at long context. An average that lands 0.8 points below GPT-6.1 Sol hides that split. For how the frontier rivals compare in general, see our GPT-6.1 Sol cost-efficiency analysis and the Opus 5.5 launch benchmarks.

Two details are worth noting. The very low scores for some general models (Gemini 4 Argon at 6.5 on BioReason Coding, Opus 5.5 at 1.1 on NextBase) suggest those tasks punish models that were never trained on raw DNA, and may also reflect prompt-format sensitivity. Second, Carbon-X per-task results on GenomeQA are uneven: 93.2% on organism group and 90.0% on motifs, but 29.8% on splice sites and 34.6% on histone marks.

The "1/1,800 the cost" claim: claimed vs. verified

The cost gap is plausible on its face. Carbon-X answers directly without extended reasoning, so each problem costs few output tokens, while the frontier models were run at their highest-scoring reasoning effort. That is a fair way to show best-case accuracy for the rivals, but it also means the ratio compares a short answer to a long chain of thought. A similar pattern appears whenever a small tuned model is set against a frontier one, as in our write-up of Shopify's post-trained Sidekick model and the VocabX specialization result. The pattern is real, but the exact multiple depends on the setup.

Limitations the authors list themselves

Two green lenses inspecting a research sheet, a picture of why Carbon-X benchmark claims need independent review

The honest part of the card is its limitations section, and it matters more than the average score. According to the card:

• It has not been clinically validated and should not be used for medical decisions.
• It can miscount bases, confuse coordinates, or make incorrect edits, especially at exact positions in long sequences.
• It may invent a sequence's identity or function when the prompt lacks context.
• Splice-site and histone-mark questions are weak.
• Quality drops on long DNA generation: it was built to read and explain sequences, not design them.
• Thinking mode exists in the chat template but was not used in post-training or evaluation.

Genetic data is also sensitive, and the card points out that privacy depends on how you access the model: local download, API, third-party provider or Space. The 14B size is part of the pitch here. Because the weights fit on a single 80 GB GPU, a lab can keep unpublished or patient-linked sequences on its own hardware, which a hosted frontier API cannot offer.

How to try Carbon-X

The model is a standard causal LM with a custom tokenizer, so you need Transformers v5 or later and trust_remote_code=True:

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")

Then pass a chat message with a reference sequence, a variant and a question, and call apply_chat_template(..., enable_thinking=False). The card recommends keeping thinking off and using the shipped sampling defaults (temperature 0.6, top-p 0.95, top-k 20), which were used for every reported result. To serve it, run vllm serve HuggingFaceBio/Carbon-X --trust-remote-code and call it through an OpenAI-compatible client.

Prompting tips from the card: put the sequence and the question in the same message, include the gene, strand, coordinates or annotations you already have, and give both reference and alternate sequences for variant questions.

GGUF and llama.cpp are not supported, which matters if you were hoping to run it on a laptop. A Hub repo named Carbon-X-GGUF exists, but the card says the custom DNA tokenizer cannot load in GGUF runtimes, so check the repo notes before relying on it.

Using it as an agent tool

The card includes recipes for calling Carbon-X from a general assistant: a generic tool-calling example, an Inspect agent example, and a ready-made skill for coding agents such as Claude Code and Codex CLI (about 80 GB of GPU memory and vLLM required). The advice is practical: pass DNA by reference rather than pasting it into the agent's context, explicitly tell the agent to use Carbon-X, do exact operations like counting and coordinate arithmetic in code, and cap generation length. That fits the broader pattern of specialist open models as tools inside agent loops, like Recursive Flint on the coding side.

How it compares to AlphaGenome and Evo 2

Google DeepMind's AlphaGenome and Arc Institute's Evo 2 are the two well-known genomics systems usually mentioned alongside this release. Carbon-X sits on a different axis: it is a conversational model that explains, which makes it easy to wire into an agent but means its answers are text, not calibrated scores. The card's stated context window is 40,960 tokens; at six bases per DNA token that nominal window is large, but inputs much longer than those seen in training have not been evaluated, so test your own sequence lengths.

The scope fits a wider push by AI labs into biology. OpenAI's GeneBench Pro work tested frontier models on computational biology, NVIDIA released an open protein design model, and Sophont's Holon-S1 predicts gene expression from pathology slides. Carbon-X's distinct angle is that it is open, small, and speaks natural language.

What this means for builders

If you build biology tooling, Carbon-X is worth an afternoon: it is Apache 2.0, runs on one GPU, and its limitations are written down. Treat it as a triage and exploration layer, such as sorting a long list of variants for a human to review, not as a diagnostic. If you evaluate models, rerun the authors' two homemade benchmarks before trusting the average, and test it on your own variants, because benchmark-built-by-the-author results are the weakest kind of evidence. If you are weighing a frontier API against a tuned 14B model for a narrow task, the card is a clean example of the tradeoff: slightly lower accuracy on harder cases, a very large cost gap, and local control of sensitive data.

What to watch next

Three things would change the picture. An independent replication on public benchmarks the authors did not write. A wet-lab or prospective clinical comparison, which Wired says does not exist yet. And a look at how Carbon-X performs when the gene name and other hints are absent, which is the real-world case for uncertain variants. Hugging Face's pace on this line, Carbon-A on October 13 and Carbon-X nine days later, suggests more releases, and we will update this post as third-party results appear.

Related reading

• Hugging Face Carbon-A: 566 million candidate genes across 22,617 species
• Holon-S1: gene expression from routine pathology slides
• OpenAI GeneBench Pro and computational biology
• NVIDIA open protein design model RFD4 Proteina
• GPT-6.1 Sol cost efficiency
• Shopify Sidekick small model vs frontier: claimed vs verified
• Official: Carbon-X model card and the HuggingFaceBio organization

Details reflect the Carbon-X model card and press coverage as of October 22, 2026; benchmark numbers are company-reported and may change.