Baichuan-M2

Released in 2025 (technical report September 2, 2025), Baichuan-M2 is a 32-billion-parameter medical large language model from Wang Xiaochuan's Baichuan Intelligence, trained with domain adaptation plus multi-stage reinforcement learning against a large verifier system that simulates interactive patient encounters instead of relying only on static benchmarks. On HealthBench it scored 60.1 — ahead of OpenAI's gpt-oss-120b (57.6) and open leaders like Qwen3-235B and Kimi K2 — and 34.7 on the harder HealthBench subset, making it one of two models worldwide to cross 32 at the time.

The design priority is deployability in privacy-sensitive clinical settings: quantized, it runs on a single RTX 4090 (about $1,400 of hardware, roughly 1/57th the cost of a DeepSeek-R1 H20 dual-node setup per the company) and is compatible with mainstream domestic Chinese chips, so hospitals can deploy on existing hardware without sending patient data to the cloud.

Limitations: it is a medical model, not a general assistant — do not use for clinical decisions without qualified oversight, and the benchmark wins are vendor-reported on their own verifier pipeline. Check the community license terms before commercial deployment; medical-device regulation may apply in your jurisdiction.

HealthBench 60.1 (vs gpt-oss-120b 57.6); HealthBench Hard 34.7 (vendor-reported)