← All models

Baichuan-M2-32B

Baichuan · 32.8B parameters

Baichuan-M2-32B is Baichuan's open-weight medical-enhanced reasoning model, which Baichuan says it released in August 2025; it was post-trained from Qwen2.5-32B with medical mid-training and multi-stage reinforcement learning against a Large Verifier System built around a patient simulator and dynamically generated clinical rubrics. The dense 32,763,876,352-parameter checkpoint keeps Qwen2.5-32B's 64-layer GQA architecture with a 131,072-position configuration, and its chat template exposes on, off, and auto thinking modes. It is released under Apache 2.0; an EAGLE-3 draft model for speculative decoding ships separately in the repository's draft folder and is not counted here.

Benchmark scores