Hugging Face has released Carbon-X, an open-source language model with 14.8 billion parameters designed for DNA sequence analysis and generation. The model is based on the Qwen3-14B architecture and utilizes a hybrid tokenizer that combines standard Byte-Pair Encoding with 6-mer DNA tokens to process genetic data within a single stream.
The model underwent mid-training on 302 billion tokens, including genomic sequences, scientific literature, and web text. Unlike closed-source models that rely on output filters, Carbon-X manages safety by excluding viral sequences from its training data. It is distributed under the Apache 2.0 license, allowing local deployment for sensitive research.
Carbon-X demonstrates significant efficiency, with developers reporting an 1800x cost advantage in inference compared to the proprietary GPT-6.1 Sol model. While the proprietary model requires extensive reasoning tokens to achieve similar performance, Carbon-X performs direct inference, making it suitable for processing large volumes of genetic variants.
The release addresses concerns regarding 'shallow guardrails' in commercial AI, which often block legitimate scientific queries. Hugging Face CEO Clément Delangue noted that such filters frequently hinder researchers while remaining ineffective against malicious actors. Carbon-X enables institutions to analyze patient data locally without relying on external cloud APIs.
The project was developed in collaboration with the Italian Institute of Genetics and Medicine (TIGEM), the University of Naples Federico II, and the Beijing Zhongguancun Academy. Quantized versions of the model are available for local hardware, including consumer GPUs with 16 GB to 24 GB of VRAM.
