Baichuan-M2

This repository contains the model presented in Baichuan-M2: Scaling Medical Capability with Large Verifier System.

🌟 Model Overview

Baichuan-M2-32B is Baichuan AI's medical-enhanced reasoning model, the second medical model released by Baichuan. Designed for real-world medical reasoning tasks, this model builds upon Qwen2.5-32B with an innovative Large Verifier System. Through domain-specific fine-tuning on real-world medical questions, it achieves breakthrough medical performance while maintaining strong general capabilities.

Model Features:

Baichuan-M2 incorporates three core technical innovations: First, through the Large Verifier System, it combines medical scenario characteristics to design a comprehensive medical verification framework, including patient simulators and multi-dimensional verification mechanisms; second, through medical domain adaptation enhancement via Mid-Training, it achieves lightweight and efficient medical domain adaptation while preserving general capabilities; finally, it employs a multi-stage reinforcement learning strategy, decomposing complex RL tasks into hierarchical training stages to progressively enhance the model's medical knowledge, reasoning, and patient interaction capabilities.

Core Highlights:

• 🏆 World's Leading Open-Source Medical Model: Outperforms all open-source models and many proprietary models on HealthBench, achieving medical capabilities closest to GPT-5
• 🧠 Doctor-Thinking Alignment: Trained on real clinical cases and patient simulators, with clinical diagnostic thinking and robust patient interaction capabilities
• ⚡ Efficient Deployment: Supports 4-bit quantization for single-RTX4090 deployment, with 58.5% higher token throughput in MTP version for single-user scenarios

📊 Performance Metrics

HealthBench Scores

General Performance

Note: AIME uses max_tokens=64k, others use 32k; temperature=0.6 for all tests.

🔧 Technical Features

📗 Technical Blog: Blog - Baichuan-M2

📑 Technical Report: Arxiv - Baichuan-M2

Large Verifier System

• Patient Simulator: Virtual patient system based on real clinical cases
• Multi-Dimensional Verification: 8 dimensions including medical accuracy, response completeness, and follow-up awareness
• Dynamic Scoring: Real-time generation of adaptive evaluation criteria for complex clinical scenarios

Medical Domain Adaptation

• Mid-Training: Medical knowledge injection while preserving general capabilities
• Reinforcement Learning: Multi-stage RL strategy optimization
• General-Specialized Balance: Carefully balanced medical, general, and mathematical composite training data

⚙️ Quick Start

# 1. load model
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("baichuan-inc/Baichuan-M2-32B", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("baichuan-inc/Baichuan-M2-32B")
# 2. Input prompt text
prompt = "Got a big swelling after a bug bite. Need help reducing it."
# 3. Encode the input text for the model
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    thinking_mode='on' # on/off/auto
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
# 4. Generate text
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=4096
)
output_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
][0].tolist()
# 5. parsing thinking content
try:
    # rindex finding 151668 ()
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("
")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("
")

print("thinking content:", thinking_content)
print("content:", content)

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.9.0 or to create an OpenAI-compatible API endpoint:

• SGLang:

  python -m sglang.launch_server --model-path baichuan-inc/Baichuan-M2-32B --reasoning-parser qwen3
  

• vLLM:

  vllm serve baichuan-inc/Baichuan-M2-32B  --reasoning-parser qwen3
  

MTP inference with SGLang

1. Replace the qwen2.py file in the sglang installation directory with draft/qwen2.py.
2. Launch sglang:

python3 -m sglang.launch_server \
--model Baichuan-M2-32B \
--speculative-algorithm EAGLE3 \
--speculative-draft-model-path Baichuan-M2-32B/draft \
--speculative-num-steps 6 \
--speculative-eagle-topk 10 \
--speculative-num-draft-tokens 32 \
--mem-fraction 0.9 \
--cuda-graph-max-bs 2 \
--reasoning-parser qwen3 \
--dtype bfloat16

⚠️ Usage Notices

1. Medical Disclaimer: For research and reference only; cannot replace professional medical diagnosis or tre