Hugging Face trains DNA model with older Qwen3 version: 14B benchmark score approaches GPT-6.1

Beating AI News Flash: Hugging Face, in collaboration with Beijing Zhongguancun Academy and Italian research institutions, has open-sourced the biological model Carbon-X, enabling researchers to analyze DNA as easily as using ChatGPT. For example, by inputting two DNA sequences—one normal and one mutated—and directly asking "Will this change cause disease?", the model can provide predictions and explanations. It can also identify regions that affect gene activity, determine DNA origin, and even complete gene sequences.

Carbon-X's foundation is not the latest model, but rather Qwen3-14B released in 2025. The team added a specialized DNA-recognition vocabulary, encoding every 6 DNA bases as one Token, and continued training with approximately 302 billion Tokens of DNA, biological text, and instruction data, enabling the model to both understand gene sequences and answer questions in natural language.

In the 5 official DNA benchmark tests, Carbon-X achieved an average score of 66.7%, approaching GPT-6.1 Sol's 67.5% and surpassing all participating open-source models. In mutation judgment within protein-coding regions and DNA sequence completion tests, it even led all participating models. Based on the team's compute cost estimates, its per-question inference cost is approximately 1/1800 of the GPT-6.1 Sol API.

The model has approximately 14.8 billion actual parameters, with open weights and local quantized versions available, allowing researchers to process sensitive genetic data on their own devices.