Quark Unveils Health AI Model with Chief Physician-Level Capabilities
Quark's Health AI Model Achieves Breakthrough in Medical Reasoning
Quark has officially published its "QuarkMed Technical Report", revealing the technical architecture behind what it claims to be the first "senior physician-level" artificial intelligence system. The release comes after the model successfully completed China's rigorous written examinations across 12 core medical disciplines.
QuarkMed Technical Report cover
Specialized Training for Medical Excellence
The health-focused large language model distinguishes itself through performance that improves with task complexity, particularly excelling in sophisticated medical reasoning scenarios. According to the report, this capability stems from three specialized data categories used during training:
- Medical documents (research papers, clinical guidelines)
- Structured medical knowledge (textbooks, reference materials)
- Real-world medical records (de-identified patient cases)
The total training corpus reached approximately 1 trillion Tokens, providing the model with unparalleled domain-specific knowledge.
Breakdown of medical data sources
Two-Stage Reinforcement Learning Approach
Quark implemented a novel dual-phase reinforcement learning (RL) system to enhance model performance:
- Phase One: Focused on improving complex scenario reasoning through large-scale medical RL
- Phase Two: Incorporated a specialized reward model evaluating outputs on:
- Honesty (factual accuracy)
- Usefulness (clinical relevance)
- Compliance (regulatory adherence)
Reward signal structure for different task types
Benchmark Dominance
The technical report includes comprehensive performance evaluations:
- Outperformed comparable models (including o3-mini) on MedQA and other international datasets
- Demonstrated increasing advantage with question difficulty on China's Physician Qualification Exam (CPQExam)
- Showed particular strength in differential diagnosis and treatment planning tasks
CPQExam performance comparison
Commitment to Open Research
In a significant move for AI medicine development, Quark announced plans to publicly release its physician examination test set. The company has made both the technical report and supplementary materials available through:
Key Points:
- First AI model to pass all 12 core Chinese medical board exams
- Specialized training with 1 trillion Tokens of medical data
- Dual-phase reinforcement learning enhances safety and accuracy
- Performance advantage grows with task complexity
- Full test set to be released for research community
