CMB
Models answer 11,200 Chinese medical exam questions and questions about 74 clinical cases. Exam answers receive exact-match scores. A model judge rates clinical answers for fluency, relevance, completeness, and medical proficiency.
The clinical track uses a different scoring method from the exam track. Test answers are public.
- Made by
- Freedom Intelligence
- Released
- August 2023
- Kind
- Medicine
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Subject
- Clinical medicine
- Language
- Chinese
- Interaction
- Single turn
- Items from
- Exams and contests
- Refresh
- One fixed set
- Run by
- Anyone
- Occupations
- Healthcare practitioners and technical
- In views
- Biology and medicine
- Last model added
- November 2023
- Scores refreshed
- November 20th, 2026
In the authors’ words
Large Language Models (LLMs) provide a possibility to make a great breakthrough in medicine. The establishment of a standardized medical benchmark becomes a fundamental cornerstone to measure progression. However, medical environments in different regions have their local characteristics, e.g., the ubiquity and significance of traditional Chinese medicine within China. Therefore, merely translating English-based medical evaluation may result in contextual incongruities to a local region. To solve the issue, we propose a localized medical benchmark called CMB, a Comprehensive Medical Benchmark in Chinese, designed and rooted entirely within the native Chinese linguistic and cultural framework. While traditional Chinese medicine is integral to this evaluation, it does not constitute its entirety. Using this benchmark, we have evaluated several prominent large-scale LLMs, including ChatGPT, GPT-4, dedicated Chinese LLMs, and LLMs specialized in the medical domain. We hope this benchmark provide first-hand experience in existing LLMs for medicine and also facilitate the widespread adoption and enhancement of medical LLMs within China. Our data and code are publicly available at https://github.com/FreedomIntelligence/CMB.
HuatuoGPT-II-34B (华佗II) tops the leaderboard with 76.8% Yi-6B-Chat leads with 66.9% among models released by November 2023
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The two tracks use different scales, so the headline is the CMB-Exam average accuracy. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Yi-6B-Chat, at 66.9%. This view shows only models we track, each at its best.
The newest model with a score here came out in November 2023, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 76.8 HuatuoGPT-II-34B (华佗II)
- 74.8 Qwen-72B-Chat
- 72.8 JianPeiGPT
- 71.2 Yi-34B-Chat
- 66.9
Yi-6B-Chat
- 64.4 ShuKunGPT
- 63.1 HuatuoGPT-II-13B (华佗)
- 60.2 GPT-4 (snapshot unspecified)
- 59.4 HuatuoGPT-II-7B (华佗)
- 55.5
DeepSeek LLM 67B
- 55.0 AntGLM-Med
- 49.6 Baichuan2-13B-Chat
- 47.2 Qwen-7B-Chat
- 43.3 Baichuan2-7B-Chat
- 42.0 DISC-MedLLM-13B (CMB-Exam-Train)
- 40.7 Qwen-1.8B-Chat
- 40.0 ChatGLM3-6B
- 40.0 Sunsimiao-7B (孙思邈)
- 39.8 IvyGPT-13B
- 39.8 DISC-MedLLM-13B
- 38.3 ChatGPT
- 38.2 internlm-chat-20b
- 36.4 Mixtral-8x7B-32kseqlen
- 34.9 Internlm-Chat-7B
- 26.1 MedicalGPT-7B
- 22.3
Mistral-7B-Instruct-v0.1
- 21.6 ChatMed-Consult-7B
- 21.6 ChatGLM-Med-7B
- 20.4 Bentsao-7B (本草)
- 7.38 BianQue-2 (扁鹊-2)
- 3.17 DoctorGLM
- 1 HuatuoGPT-II-34B (华佗II) Zero-shot, answer only 76.8%
- 2 Qwen-72B-Chat Three-shot, answer only 74.8%
- 3 JianPeiGPT Best reported, generation strategy unspecified 72.8%
- 4 Yi-34B-Chat Three-shot, answer only 71.2%
- 5
Yi-6B-Chat Three-shot, answer only 66.9%
- 6 ShuKunGPT Best reported, generation strategy unspecified 64.4%
- 7 HuatuoGPT-II-13B (华佗) Three-shot, answer only 63.1%
- 8 GPT-4 (snapshot unspecified) Three-shot, answer only 60.2%
- 9 HuatuoGPT-II-7B (华佗) Three-shot, answer only 59.4%
- 10
DeepSeek LLM 67B Three-shot, answer only 55.5%
- 11 AntGLM-Med Best reported, generation strategy unspecified 55%
- 12 Baichuan2-13B-Chat Three-shot, answer only 49.6%
- 13 Qwen-7B-Chat Three-shot, answer only 47.2%
- 14 Baichuan2-7B-Chat Zero-shot, answer only 43.3%
- 15 DISC-MedLLM-13B (CMB-Exam-Train) Three-shot, answer only 42%
- 16 Qwen-1.8B-Chat Three-shot, answer only 40.7%
- 17 ChatGLM3-6B Zero-shot, answer only 40%
- 18 Sunsimiao-7B (孙思邈) Three-shot, chain of thought 40%
- 19 IvyGPT-13B Three-shot, chain of thought 39.8%
- 20 DISC-MedLLM-13B Zero-shot, chain of thought 39.8%
- 21 ChatGPT Zero-shot, answer only 38.3%
- 22 internlm-chat-20b Zero-shot, chain of thought 38.2%
- 23 Mixtral-8x7B-32kseqlen Three-shot, answer only 36.4%
- 24 Internlm-Chat-7B Zero-shot, answer only 34.9%
- 25 MedicalGPT-7B Three-shot, answer only 26.1%
- 26
Mistral-7B-Instruct-v0.1 Zero-shot, answer only 22.3%
- 27 ChatMed-Consult-7B Three-shot, chain of thought 21.6%
- 28 ChatGLM-Med-7B Three-shot, answer only 21.6%
- 29 Bentsao-7B (本草) Three-shot, answer only 20.4%
- 30 BianQue-2 (扁鹊-2) Three-shot, chain of thought 7.38%
- 31 DoctorGLM Zero-shot, answer only 3.17%
Source: The CMB paper
Source: The CMB paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | HuatuoGPT-II-34B (华佗II) open weights | 76.8% | Zero-shot, answer only | Official leaderboard | |
| 2 | Qwen-72B-Chat open weights | 74.8% | Three-shot, answer only | Official leaderboard | |
| 3 | JianPeiGPT closed weights | 72.8% | Best reported, generation strategy unspecified | Official leaderboard | |
| 4 | Yi-34B-Chat open weights | 71.2% | Three-shot, answer only | Official leaderboard | |
| 5 | | 66.9% | Three-shot, answer only | Nov 2023 | Official leaderboard |
| 6 | ShuKunGPT closed weights | 64.4% | Best reported, generation strategy unspecified | Official leaderboard | |
| 7 | HuatuoGPT-II-13B (华佗) open weights | 63.1% | Three-shot, answer only | Official leaderboard | |
| 8 | GPT-4 (snapshot unspecified) closed weights | 60.2% | Three-shot, answer only | Official leaderboard | |
| 9 | HuatuoGPT-II-7B (华佗) open weights | 59.4% | Three-shot, answer only | Official leaderboard | |
| 10 | | 55.5% | Three-shot, answer only | Nov 2023 | Official leaderboard |
| 11 | AntGLM-Med closed weights | 55% | Best reported, generation strategy unspecified | Official leaderboard | |
| 12 | Baichuan2-13B-Chat | 49.6% | Three-shot, answer only | Sep 2023 | Official leaderboard |
| 13 | Qwen-7B-Chat open weights | 47.2% | Three-shot, answer only | Official leaderboard | |
| 14 | Baichuan2-7B-Chat open weights | 43.3% | Zero-shot, answer only | Official leaderboard | |
| 15 | DISC-MedLLM-13B (CMB-Exam-Train) open weights | 42% | Three-shot, answer only | Official leaderboard | |
| 16 | Qwen-1.8B-Chat open weights | 40.7% | Three-shot, answer only | Official leaderboard | |
| 17 | ChatGLM3-6B | 40% | Zero-shot, answer only | Oct 2023 | Official leaderboard |
| 18 | Sunsimiao-7B (孙思邈) open weights | 40% | Three-shot, chain of thought | Official leaderboard | |
| 19 | IvyGPT-13B open weights | 39.8% | Three-shot, chain of thought | Official leaderboard | |
| 20 | DISC-MedLLM-13B open weights | 39.8% | Zero-shot, chain of thought | Official leaderboard | |
| 21 | ChatGPT closed weights | 38.3% | Zero-shot, answer only | Official leaderboard | |
| 22 | internlm-chat-20b | 38.2% | Zero-shot, chain of thought | Sep 2023 | Official leaderboard |
| 23 | Mixtral-8x7B-32kseqlen open weights | 36.4% | Three-shot, answer only | Official leaderboard | |
| 24 | Internlm-Chat-7B open weights | 34.9% | Zero-shot, answer only | Official leaderboard | |
| 25 | MedicalGPT-7B open weights | 26.1% | Three-shot, answer only | Official leaderboard | |
| 26 | | 22.3% | Zero-shot, answer only | Sep 2023 | Official leaderboard |
| 27 | ChatMed-Consult-7B open weights | 21.6% | Three-shot, chain of thought | Official leaderboard | |
| 28 | ChatGLM-Med-7B open weights | 21.6% | Three-shot, answer only | Official leaderboard | |
| 29 | Bentsao-7B (本草) open weights | 20.4% | Three-shot, answer only | Official leaderboard | |
| 30 | BianQue-2 (扁鹊-2) open weights | 7.38% | Three-shot, chain of thought | Official leaderboard | |
| 31 | DoctorGLM open weights | 3.17% | Zero-shot, answer only | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 66.9% | Three-shot, answer only | Nov 2023 | Official leaderboard |
| 2 | | 55.5% | Three-shot, answer only | Nov 2023 | Official leaderboard |
| 3 | Baichuan2-13B-Chat | 49.6% | Three-shot, answer only | Sep 2023 | Official leaderboard |
| 4 | ChatGLM3-6B | 40% | Zero-shot, answer only | Oct 2023 | Official leaderboard |
| 5 | internlm-chat-20b | 38.2% | Zero-shot, chain of thought | Sep 2023 | Official leaderboard |
| 6 | | 22.3% | Zero-shot, answer only | Sep 2023 | Official leaderboard |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The CMB paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Three-shot, answer only | 66.9% | ||||
| Zero-shot, answer only | 65.9% | ||||
| Zero-shot, chain of thought | 63% | ||||
| Three-shot, chain of thought | 58.2% | ||||
| Three-shot, answer only | 55.5% | ||||
| Zero-shot, chain of thought | 52.3% | ||||
| Zero-shot, answer only | 52% | ||||
| Three-shot, chain of thought | 50.6% | ||||
| Baichuan2-13B-Chat | Three-shot, answer only | 49.6% | |||
| Baichuan2-13B-Chat | Zero-shot, answer only | 48.9% | |||
| Baichuan2-13B-Chat | Zero-shot, chain of thought | 48.4% | |||
| Baichuan2-13B-Chat | Three-shot, chain of thought | 46.7% | |||
| ChatGLM3-6B | Zero-shot, answer only | 40% | |||
| ChatGLM3-6B | Three-shot, answer only | 38.7% | |||
| internlm-chat-20b | Zero-shot, chain of thought | 38.2% | |||
| internlm-chat-20b | Zero-shot, answer only | 38.2% | |||
| ChatGLM3-6B | Zero-shot, chain of thought | 37.8% | |||
| internlm-chat-20b | Three-shot, answer only | 37.5% | |||
| ChatGLM3-6B | Three-shot, chain of thought | 35.8% | |||
| internlm-chat-20b | Three-shot, chain of thought | 33.5% | |||
| Zero-shot, answer only | 22.3% | ||||
| Three-shot, answer only | 19.9% | ||||
| Zero-shot, chain of thought | 19% | ||||
| Three-shot, chain of thought | 19% |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The CMB paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Three-shot, answer only | 66.9% | ||||
| Zero-shot, answer only | 65.9% | ||||
| Zero-shot, chain of thought | 63% | ||||
| Three-shot, chain of thought | 58.2% | ||||
| Three-shot, answer only | 55.5% | ||||
| Zero-shot, chain of thought | 52.3% | ||||
| Zero-shot, answer only | 52% | ||||
| Three-shot, chain of thought | 50.6% | ||||
| Baichuan2-13B-Chat | Three-shot, answer only | 49.6% | |||
| Baichuan2-13B-Chat | Zero-shot, answer only | 48.9% | |||
| Baichuan2-13B-Chat | Zero-shot, chain of thought | 48.4% | |||
| Baichuan2-13B-Chat | Three-shot, chain of thought | 46.7% | |||
| ChatGLM3-6B | Zero-shot, answer only | 40% | |||
| ChatGLM3-6B | Three-shot, answer only | 38.7% | |||
| internlm-chat-20b | Zero-shot, chain of thought | 38.2% | |||
| internlm-chat-20b | Zero-shot, answer only | 38.2% | |||
| ChatGLM3-6B | Zero-shot, chain of thought | 37.8% | |||
| internlm-chat-20b | Three-shot, answer only | 37.5% | |||
| ChatGLM3-6B | Three-shot, chain of thought | 35.8% | |||
| internlm-chat-20b | Three-shot, chain of thought | 33.5% | |||
| Zero-shot, answer only | 22.3% | ||||
| Three-shot, answer only | 19.9% | ||||
| Zero-shot, chain of thought | 19% | ||||
| Three-shot, chain of thought | 19% |
2 models have held number one in the benchmark’s paper since its release in August 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The CMB paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official CMB leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Law, finance and medicine
- MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy.
Qwen 3.8 Max 80.4%
- HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
- HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
- FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy.
Ling-3.1-flash 81.7%
- KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. o1-preview 92.7%
- Harvey LAB-AA v1.1 In use Artificial Analysis's run of Harvey's Legal Agent Benchmark on Harvey's newer private set of 120 tasks across 24 practice areas. A deliverable with a material hallucination scores zero, and the headline is the share of judges who find every rubric criterion passed. GPT-6.2 Astra 15%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 3 years ago, in November 2023, for DeepSeek LLM 67B. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (31 entries).
- Pass Models scored 6 models we track, plus 25 other entries on its leaderboard.
- Warning Newest score The newest came 3 years ago, in November 2023, for DeepSeek LLM 67B.
- Note Frontier models None of its models are on the capabilities index.
- Pass Headroom The best is 66.9%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 24 rows for models we track and 25 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.