Robots Atlas>ROBOTS ATLAS
Carbon-X
AI Models›Carbon

Carbon-X

X · Family: Carbon
Open 14.8B language model for DNA understanding and generation from the Carbon family; predicts mutation pathogenicity and analyses genomic sequences.
✓ Active✓ Public access⚖ Open sourceLLMScientific AI📁 Carbon
Context window
40 960
tokens
Parameters
14.8B
parameters
Release date
23 October 2026
Access:DownloadDeployment:💻 Local☁ Cloud

Overview

Carbon-X is an open 14.8B-parameter decoder-only language model from the Carbon family, developed by the Hugging Face Bio team. It was trained jointly on DNA sequences and natural-language text, starting from Qwen3-14B.

DNA and text share a single token stream in a single decoder - there is no separate DNA encoder. A hybrid tokenizer keeps a BPE vocabulary for text and encodes DNA as non-overlapping 6-mers, so one DNA token covers 6 base pairs.

You give the model raw DNA inside an ordinary chat conversation and it 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.

Across five DNA benchmarks Carbon-X averages 66.7% with thinking disabled - 27 points above the base Qwen3-14B. The weights fit on a single 80 GB GPU in bfloat16, so researchers can keep sensitive or unpublished sequences on their own infrastructure. The model is released under the Apache 2.0 license.

Classification
LLMScientific AI
Family: Carbon
Applications
GenomicsBioinformaticsComputational biology
Access & deployment
Download
LocalCloud
Weights: Open source
Key parameters
📏 Context: 40 960
🧩 Parameters: 14.8B
✓ Tools · ✓ Fine-tuning
📥 Input: text

Technical specification

Context window
40 960
tokens
Parameters
14.8B
parameters
License
Apache 2.0
Hardware requirements
Weights fit on a single 80 GB GPU in bfloat16. Serves with vLLM on a single GPU.
Features:✓ Tool use✓ Fine-tuning
Modalities
⬇ Input
text
⬆ Output
text

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to perform multi-step logical inference, solve complex problems and decompose tasks into steps.
Category: reasoning
Long context
Processing very long inputs (tens to hundreds of thousands of tokens) while maintaining coherence.
Category: language
Variant effect prediction
Predicting whether a mutation in a DNA sequence is benign or pathogenic from the sequence alone.
Category: other
DNA sequence understanding
Classifying and interpreting features of DNA sequences: regulatory regions, splice sites, organism group, motifs and transcription-factor binding sites.
Category: other
Next-base prediction
Predicting the next bases (nucleotides) of a DNA sequence.
Category: other
Application domains
GenomicsBioinformaticsComputational biology

Benchmark results

6 benchmarks
BioReason Coding
balanced accuracy · Thinking disabled, default sampling, one sample per problem.
74.9%
📄 Karta modelu Carbon-X (Hugging Face)
BioReason Non-SNV
balanced accuracy · Thinking disabled, default sampling, one sample per problem.
85.1%
📄 Karta modelu Carbon-X (Hugging Face)
GenomeQA
accuracy · Thinking disabled, default sampling, one sample per problem.
63.6%
📄 Karta modelu Carbon-X (Hugging Face)
NextBase
per-base accuracy · Thinking disabled, default sampling, one sample per problem.
39.4%
📄 Karta modelu Carbon-X (Hugging Face)
FrontierVEP
accuracy · Thinking disabled, default sampling, one sample per problem.
70.4%
📄 Karta modelu Carbon-X (Hugging Face)
IFEval
prompt-level strict accuracy · Thinking disabled, default sampling, one sample per problem.
79.3%
📄 Karta modelu Carbon-X (Hugging Face)

Technical architecture

Sources and related pages