Carbon-X: 14.8B Open-Source LLM for Analyzing DNA Sequences via Chat

Carbon-X is a 14.8B parameter open-source language model trained simultaneously on DNA sequences and natural language text. When raw DNA data is input into a standard chat interface, it responds in English, predicting whether mutations are harmful or benign and identifying regions that regulate gene activity.

1 / 2

Unlike existing methods, it processes data as a single token stream without separating the DNA encoder and text decoder. It is fine-tuned based on Qwen3-14B and released under the Apache 2.0 license. The model performs tasks ranging from next-base prediction to organism group classification and explanation of reasoning processes, with responses potentially mixing DNA and natural language.

It is compatible with various inference servers such as vLLM, SGLang, and Docker Model Runner, allowing for easy deployment via an OpenAI-compatible API. It can be used as an agent tool to handle complex biological questions or as a direct text-based analysis tool for variant effect prediction research.