Carbon-X
An open-source 14.8B-parameter decoder-only model jointly trained on DNA and text. DNA is tokenized with 6-mers and shares a vocabulary with natural language, allowing DNA sequences to be entered directly in a chat interface; the model responds in English. Training on both modalities within a single model enables the handling of biological sequences and natural-language dialogue in one system.
Timeline (last 90 days)
Hugging Face released an open-source DNA language model named Carbon-X.
Carbon-X has 14.8B parameters.
Carbon-X is built on Qwen3-14B.
Carbon-X uses the Apache 2.0 license.
According to MarkTechPost, Carbon-X is the first to connect DNA and natural language within the same model.
The sources, scale and cleaning methods of Carbon-X's pretraining corpus were not disclosed in the original material.
The specific composition of Carbon-X's 5 benchmarks has not yet been made public.
In DNA functional annotation scenarios, the model may give specific regulatory function or disease association descriptions for unknown sequences, and these descriptions have no experimental evidence to support them, posing high risk.
Carbon-X is a 14.8B-parameter, decoder-only, Apache 2.0-licensed open-source model that handles both DNA and natural language, with DNA split into 6-mers and sharing a vocabulary with text, and mid-trained on 302B tokens.
Carbon-X accepts raw DNA sequence text in the conversation, 6-mer tokenization is done internally, and file parsing and format conversion are best handled by external scripts.
7 more timeline events
All Facts (17)
In DNA functional annotation scenarios, the model may give specific regulatory function or disease association descriptions for unknown sequences, and these descriptions have no experimental evidence to support them, posing high risk.
50%UnverifiedCarbon-X has 14.8B parameters.
50%UnverifiedCarbon-X is built on Qwen3-14B.
50%UnverifiedCarbon-X uses the Apache 2.0 license.
50%UnverifiedAccording to MarkTechPost, Carbon-X is the first to connect DNA and natural language within the same model.
50%UnverifiedThe sources, scale and cleaning methods of Carbon-X's pretraining corpus were not disclosed in the original material.
50%UnverifiedThe specific composition of Carbon-X's 5 benchmarks has not yet been made public.
50%UnverifiedThe comparability of Carbon-X's benchmark comparison with GPT-6.1 Sol is limited, because the evaluation details, prompt design, and toolchain enablement of the closed-source model are not transparent.
50%UnverifiedCarbon-X is a 14.8B-parameter, decoder-only, Apache 2.0-licensed open-source model that handles both DNA and natural language, with DNA split into 6-mers and sharing a vocabulary with text, and mid-trained on 302B tokens.
50%UnverifiedCarbon-X accepts raw DNA sequence text in the conversation, 6-mer tokenization is done internally, and file parsing and format conversion are best handled by external scripts.
50%UnverifiedWith thinking mode disabled, Carbon-X scores an average of 66.7% across five DNA benchmarks, compared with 67.5% for GPT-6.1 Sol.
50%UnverifiedCarbon-X's weights fit into a single 80GB GPU at bfloat16 precision.
50%UnverifiedThe model card explicitly states that Carbon-X has not undergone clinical validation and that precise sequence operations should still be performed by code.
50%UnverifiedCarbon-X is suited to bioinformatics teams processing sensitive genomic data on intranet, developers integrating DNA analysis into agent workflows, and researchers of genomic foundation models.
50%UnverifiedCarbon-X is not recommended for use as a clinical decision-making tool or as a replacement for conventional alignment and variant annotation pipelines.
50%UnverifiedImprovements on public benchmarks do not necessarily hold on private variant datasets, because real-world data suffer from issues such as batch effects, uneven sequencing depth, and population specificity.
50%UnverifiedHugging Face released an open-source DNA language model named Carbon-X.
50%