Baichuan-M2

What it is

Baichuan-M2 is an open-source medical-enhanced model from Baichuan, the Chinese AI company. Its full name in the model repository is Baichuan-M2-32B, a 32-billion-parameter model that builds on Qwen2.5-32B. ^[1][2] It was released under Apache 2.0 for research and reference, and Baichuan stated that it is not a substitute for medical diagnosis. ^[1]

Origins

Baichuan-M2 was unveiled in August 2025 as an open-source model designed for private deployment in clinical settings. ^[4] It is the company's second open-source medical-enhanced model, following the first in its medical line. ^[1]

Key features

Baichuan called Baichuan-M2 the world's strongest open-source medical model, and said it was the open model closest to GPT-5 in medical capability. ^[1] On HealthBench, a medical evaluation benchmark, it scored 60.1 overall, 34.7 on HealthBench-Hard and 91.5 on HealthBench-Consensus. For comparison, the same benchmark figures were 57.6, 30.0 and 90.0 for gpt-oss-120b, and 53.6, 22.6 and 91.5 for DeepSeek-R1-0528. ^[1]

Training used what Baichuan calls a Large Verifier System. It includes a patient simulator built on real clinical cases, verification across eight dimensions, medical mid-training and multi-stage reinforcement learning. ^[2] Reinforcement learning means the model improves by trying answers and receiving feedback on them.

On general benchmarks the model scored 83.4 on AIME24 and 72.9 on AIME25, which are competition-level maths tests. Baichuan compared these with 81.4 and 72.9 for Qwen3-32B in thinking mode. ^[1] The model supports thinking mode switched on, off or set to automatic. ^[1]

With 4-bit quantisation the model runs on a single RTX 4090 graphics card. Baichuan also released an MTP version, which it said raises single-user throughput by 58.5%. ^[2] A W8A8 version for Huawei Ascend chips was published as well. ^[1]

Licence and purpose

The model is released under Apache 2.0, and Baichuan describes it as intended for research and reference, not as a replacement for medical diagnosis. ^[1]

Position in development

The technical report, arXiv:2509.02208, was published on 2 September 2025. ^[3][2] Its scores are the company's own figures, so the technical report is the place to check the test conditions.