Baichuan-4
by Baichuan AI·China·Released
Baichuan AI's flagship — strong Chinese-language and medical-domain performance.
About this model
Baichuan-4 (May 2024) is Baichuan AI's flagship closed-weights model — founder Wang Xiaochuan's bet that domain specialisation (medical, legal) is more valuable than chasing general benchmarks. Baichuan-M1, the medical-specialist variant on the same backbone, has been deployed in several Chinese hospital pilots.
On general benchmarks Baichuan-4 sits behind Qwen/DeepSeek/Kimi but has a meaningful lead on Chinese-medical evaluation suites. The lab also runs an open-weights track (Baichuan2-7B/13B) for research use.
Strengths
- •Specialised medical / legal performance
- •Strong Chinese-language fundamentals (C-Eval 88.4%)
- •Open-weights Baichuan2 variants for research
Limitations
- •Smaller scale than peer Chinese frontier labs
- •Closed weights for the flagship Baichuan-4 itself
- •International developer adoption minimal
When to use it
- →Chinese medical AI deployments (hospital pilots)
- →Legal / regulatory document analysis in Chinese
- →Research collaboration via open Baichuan2 variants
Architecture & training
Architecture undisclosed beyond 'dense transformer family'. Baichuan AI has published the Baichuan2 technical report which documents the open variants in detail.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| C-Eval | 88.4 |