← All models

Baichuan-M3-235B

Baichuan · 235.1B parameters

Baichuan-M3-235B is Baichuan's open-weight medical-enhanced model fine-tuned from Qwen3-235B-A22B and trained to model the clinical decision process of history taking, differential diagnosis, laboratory testing, and final diagnosis, using segmented-pipeline reinforcement learning (SPAR) and fact-aware reinforcement learning to curb hallucinations. The checkpoint keeps Qwen3's 94-layer, 128-expert MoE architecture with eight experts active per token and contains 235,093,634,560 parameters, excluding a separately shipped Gated EAGLE-3 draft model; its released configuration supports 40,960 positions. Its chat template always opens a thinking block, and the weights are released under Apache 2.0.

Benchmark scores