Recently, Quark officially released the health large model technical report "QuarkMed Technical Report," revealing for the first time the technical implementation details of the "senior physician-level" capabilities.
QuarkMed Technical Report
Previously, the Quark Health Large Model successfully passed the written exam evaluation of 12 core medical disciplines in China, becoming the first large model to complete this challenge domestically. Compared with general models, the Quark Health Large Model shows a performance curve where the higher the difficulty, the more obvious the advantage, especially achieving breakthroughs in complex medical reasoning tasks. This technical report systematically discloses the key paths and technical highlights behind this breakthrough.
Face with the need for high-quality, highly professional training data for medical models, the Quark Health Large Model used three types of core medical data at different stages of model training: medical documents, medical knowledge, and medical records, with a total data volume of about 1 trillion Tokens. These professional data can effectively compensate for the shortcomings of pre-training corpora, helping to improve the accuracy and reasoning ability of the model.
Classification and scale of medical data sources
To enhance the model's correctness, safety, and complex reasoning ability, the Quark Health Large Model introduced two reinforcement learning (RL) stages. The first stage improves the model's reasoning ability in complex scenarios through large-scale medical reinforcement learning. The second stage designs a reward model to evaluate the quality of model output from three perspectives: honesty, usefulness, and content compliance, adjusting the model's behavior to align with human preferences and values.
Three types of reward signals used during training for general tasks and reasoning tasks
The technical report also公布了 multiple performance test results. In international authoritative datasets such as MedQA, the Quark Health Large Model demonstrated superior performance compared to same-size models like o3-mini. In the Chinese Physician Qualification Exam (CPQExam) written evaluation, the higher the difficulty, the more obvious the advantage of the Quark Health Large Model.
CPQExam test results
The report revealed that Quark plans to fully open the physician examination test set to promote AI research related to medicine.
Access and download addresses are as follows:
https://arxiv.org/pdf/2508.11894
https://github.com/Quark-Medical/QuarkMed/blob/main/report/QuarkMed_Technical_Report.pdf
