An exam written by academicians has picked China's top medical LLM. Per Red Star News and Sina Tech, the latest MedBench rankings (2026.10.08) released on October 8 place the self-developed "Huashu Zhiyi" medical LLM from Chengdu-based Huaxi Digital Medicine (HDM) first nationally with a composite score of 76.3, with another HDM model at No.3. The next day, its underlying "Medical AI Omni-Intelligent Platform" debuted at the first China–Central Asia maternal and child health forum in Kazakhstan.

How hard is an exam written by academicians?

MedBench, led by Shanghai AI Laboratory and first launched in 2024, has iterated to version 5.0 with 63 fine-grained clinical evaluation tasks and nearly 1 million questions. Teams led by academicians Zhong Nanshan, Ge Junbo and Fan Xianqun head the respiratory, cardiology and ophthalmology tracks; West China Hospital and 10+ top hospitals co-build it. It is China's first native medical vertical benchmark and a designated pre-deployment evaluation vehicle. Baidu's Lingyi, JD's Jingyi Qianxun and Google's Gemini-3.1-Pro-Preview also rank near the top.

Key numbers

· MedBench composite: 76.3, national No.1 (another model at No.3)

· Pre-consultation intake time: 5.4 → 4.4 minutes, down ~18.5%

· Task completion: 64.5% → 71.3% (+6.8 pp)

· Information completeness: 85.6% → 93.1% (+7.5 pp)

· Deployed in multiple key public hospitals, serving 700,000+ patients cumulatively

How the top score was built

Architecture is three-layer: MDAM-series foundation models at the base; an AIOS platform layer managing agents and data; and applications for patients, clinicians and hospital management on top. Hospitals need no self-built compute — pay per token — or can deploy on-premise with data never leaving the hospital; a public-facing "Huahua Assistant" offers 24/7 consultation and accompaniment. Founded in 2023 by West China Second University Hospital with Wuhou state capital, HDM raised tens of millions of RMB in a 2025 Pre-A round.

Numbers from the clinic

Beyond rankings, harder evidence comes from a published paper: in OB-GYN and pediatrics pilots at West China Second Hospital, intake time fell ~18.5%, task completion rose 6.8 points and completeness 7.5 points versus a general-purpose baseline. For a 70-year-old follow-up patient, the assistant interviews in everyday language, drafts a pre-visit summary, guides in-hospital navigation, and sends follow-up reminders — clinical judgment stays with the physicians.

The benchmark shows model capability; whether patients take fewer wrong turns is the real test of product value.

The NineZenith view

Healthcare is where industry AI is most tested on being "grounded": pre-visit summaries and follow-up education look light, yet decide whether clinicians adopt the tools. This mirrors how NineZenith's "Tianxing" industry LLMs land in government and enterprise scenarios — not chasing general leaderboards, but making one workflow smooth and auditable before scaling down to county-level care. (Information synthesized from Red Star News, Sina Tech and Sohu Finance reports)