Baichuan-M3 is Baichuan AI's new-generation medical-enhanced large language model, a major milestone following Baichuan-M2.
In contrast to prior approaches that primarily focus on static question answering or superficial role-playing, Baichuan-M3 is trained to explicitly model the clinical decision-making process, aiming to improve usability and reliability in real-world medical practice. Rather than merely producing "plausible-sounding answers" or high-frequency vague recommendations like "you should see a doctor soon," the model is trained to proactively acquire critical clinical information, construct coherent medical reasoning pathways, and systematically constrain hallucination-prone behaviors.
Core Highlights
🏆 Surpasses GPT-5.2: Outperforms OpenAI's latest model across HealthBench, HealthBench-Hard, hallucination evaluation, and SCAN-bench, establishing a new SOTA in medical AI
🩺 High-Fidelity Clinical Inquiry: The only model to rank first across all three SCAN-bench dimensions—Clinical Inquiry, Laboratory Testing, and Diagnosis
🧠 Low Hallucination, High Reliability: Achieves lower hallucination rate than GPT-5.2 through Fact-Aware RL, even without external tools
⚡ Efficient Deployment: W4 quantization reduces memory to 26% of original; Gated Eagle3 speculative decoding achieves 96% speedup
📊 Performance
HealthBench & Hallucination Evaluation
HealthBench is OpenAI's authoritative medical benchmark, constructed by 262 practicing physicians from 60 countries, comprising 5,000 high-fidelity multi-turn clinical conversations.
HealthBench Performance
Compared to Baichuan-M2, Baichuan-M3 improves by 28 percentage points on HealthBench-Hard, reaching 44.4 and surpassing GPT-5.2. It also ranks first on the HealthBench Total leaderboard.
For hallucination evaluation, we decompose long-form responses into fine-grained, verifiable atomic medical claims and validate each against authoritative medical evidence. Even in a tool-free setting, Baichuan-M3 achieves lower hallucination rate than GPT-5.2.
SCAN-bench Evaluation
SCAN-bench is our end-to-end clinical decision-making benchmark that simulates the complete clinical workflow from patient encounter to final diagnosis, evaluating models' high-fidelity clinical inquiry capabilities through three stations: History Taking, Ancillary Investigations, and Final Diagnosis.
SCAN-bench Performance
Baichuan-M3 ranks first across all three core dimensions, outperforming the second-best model by 12.4 points in Clinical Inquiry.
📢 The SCAN-bench will be open-sourced soon. Stay tuned.
🔬 Technical Features
📖 For detailed technical information, please refer to: Tech Blog
SPAR: Segmented Pipeline Reinforcement Learning
To address reward sparsity and credit assignment challenges in long clinical interactions, we propose SPAR (Step-Penalized Advantage with Relative baseline): it decomposes clinical workflows into four stages—history taking, differential diagnosis, laboratory testing, and final diagnosis—each with independent rewards, combined with process-level rewards for precise credit assignment, driving the model to construct auditable and complete decision logic.
Fact-Aware Reinforcement Learning
By integrating factual verification directly into the RL loop, we build an online hallucination detection module that validates model-generated medical claims against authoritative medical evidence in real-time, supported by efficient caching mechanisms for online RL training. A dynamic reward aggregation strategy adaptively balances task learning and factual constraints based on the model's capability stage, significantly enhancing medical factual reliability without sacrificing reasoning depth.
Efficient Training and Inference
Adopts a three-stage multi-expert fusion training paradigm (Domain-Specific RL → Offline Distillation → MOPD), combined with Gated Eagle3 speculative decoding (96% speedup) and W4 quantization (only 26% memory) for efficient deployment.
🔧 Quick Start
python
1from transformers import AutoTokenizer, AutoModelForCausalLM
23model = AutoModelForCausalLM.from_pretrained("baichuan-inc/Baichuan-M3-235B", trust_remote_code=True)4tokenizer = AutoTokenizer.from_pretrained("baichuan-inc/Baichuan-M3-235B")56messages =[{"role":"user","content":"I've been having headaches lately, especially worse in the afternoon. What should I do?"}]7text = tokenizer.apply_chat_template(8 messages,9 tokenize=False,10 add_generation_prompt=True,11 thinking_mode='on'12)13model_inputs = tokenizer([text], return_tensors="pt").to(model.device)1415generated_ids = model.generate(16**model_inputs,17 max_new_tokens=32768,18 temperature=0.619)20response = tokenizer.decode(generated_ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True)21print(response)
Deployment
Create an OpenAI-compatible API endpoint using sglang>=0.4.6.post1 or vllm>=0.9.0:
To support the gated attention in eagle3 draft model, just replace the code llama_eagle3.py in the sglang installation directory with our draft/llama_eagle3.py, for example:
