Baichuan and Tsinghua launch medical AI model Baichuan-M4, top in multiple tests

Baichuan and Tsinghua University have launched a new medical AI model, Baichuan-M4. It ranks first worldwide on the HealthBench test and its Hard and Professional subsets, scoring 68.6 overall, beating GPT-5.5 by over 10 points. Its hallucination rate is only 3.3%.

M4 can actively ask about symptoms and their causes, focusing on serious conditions first instead of waiting for complete user input. Baichuan worked with over 150 doctors to create a dynamic evaluation system based on the OSCE method used in medical education. M4 scored 79.0 on first diagnosis and 74.7 on follow-ups, both ahead of other models.

M4 also features full-course memory, linking medical history and multiple consultations, scoring 86.9 in long-context clinical memory tests. Baichuan introduced evidence anchoring, ensuring every medical conclusion is backed by specific parts of original papers or guidelines, achieving a citation accuracy of 90.0, far above GPT-5.5's 54.7.