Quark Health Large Model Passes 12 Chief Physician Exams, Once Again Raising the Limit}

On July 23, news broke that Quark Health’s large model has successfully passed the written exams for all 12 core medical disciplines required for chief physicians in China, making it the first large model in the country to achieve this milestone. Currently, the "Chief-Level AI Doctor" capability has been fully integrated into Quark’s AI search system, allowing users to perform in-depth health queries by simply selecting the deep search option.

This marks another leap in Quark’s AI capabilities after passing the associate chief physician exam in May. Compared with specialized and general models, Quark’s health model shows increasingly prominent performance advantages as the difficulty rises, demonstrating breakthroughs in complex medical reasoning tasks.

It reveals the enormous potential for developing specialized medical models. Based on the comprehensive Qianwen model, Quark Health has charted a deep engineering route tailored for vertical scenarios. As Xu Jian, head of Quark’s health algorithms, states, "We are not training AI to answer medical questions but to learn medical thinking."

One of the core breakthroughs of Quark’s health model is its development of "slow thinking" ability. This combines chain reasoning with multi-stage clinical deduction path modeling, enabling the model to derive answers step-by-step and in layers when facing complex medical issues.

The foundation of slow thinking is high-quality reasoning training data. To achieve this, Quark has built a "dual data pipeline + dual reward mechanism" system. It classifies medical data into "verifiable" and "non-verifiable" categories, corresponding to diagnostic tasks and health advice tasks, respectively. The training process introduces "process reward models" and "result reward models" to evaluate the reasoning chain’s rationality and the accuracy of final conclusions, significantly enhancing clinical interpretability and reasoning consistency.

This system also incorporates multi-stage reinforcement learning, including strict manual verification of initial data, multiple rounds of sample filtering, progressive difficulty training strategies, and mechanisms to prevent "high-score speculation" or cheating. Driven by real doctor annotations and a "question—thinking—answer" data set, Quark’s model not only learns medical knowledge but also masters pathways of medical reasoning, evidence integration, and multi-solution balancing. An authoritative medical knowledge base underpins the model’s professional and timely outputs.

Dr. Xie Jinsing, chief physician of Cardiac Surgery at Anzhen Hospital, believes that Quark’s answers in some cases are even more professional than those of experienced doctors. Behind this achievement is the deep involvement of a professional medical team. Currently, Quark’s health model is supported by a team of nearly 1,000 medical experts, over 400 of whom are associate chief physicians or higher.

Thanks to its professionalism in medicine, Quark AI search has attracted a large group of medical students and doctors. Zhao Cunzong, head of Quark’s operations, states that the platform’s monthly active users among medical students nationwide have exceeded 2 million, covering more than half of the student population. They widely use Quark for basic knowledge searches, exam preparation, and clinical decision support.