1. Abstract
Baichuan-M3-235B is a medical-enhanced large language model released by Baichuan Intelligence, emphasizing the "clinical decision-making process" as the training goal: the model not only answers questions, but also actively asks for key medical history information, organizes differential diagnosis ideas, and tries to restrain unreliable medical assertions in generation. The official announced the results of HealthBench, HealthBench-Hard, Hallucination Evaluation and Self-built SCAN-bench in the model card, and claimed to be leading in these evaluations.
2. Core features
- Dialogue strategy for clinical process: link organization output around medical history collection→ differential diagnosis→ examination recommendations→ and final diagnosis.
- SPAR segmented assembly line reinforcement learning: divide the long-link consultation into multiple stages to give rewards to alleviate the sparse rewards and credit allocation problems of long-term conversations.
- Fact-Aware RL: Integrate fact-checking into the reinforcement learning loop and impose constraints on medical "verifiable assertions" to reduce the risk of hallucinations.
- Efficient deployment: Officials provide W4 quantization and Eagle3-based speculative decoding solutions to reduce memory usage and increase throughput.
3. Installation
- Basic dependencies: Use Transformers to load (need to enable trust_remote_code) and prepare a multi-card environment that can carry 235B MoE models.
- Inference service: Officials recommend launching OpenAI-compatible APIs with vLLM or SGLang and using qwen3's reasoning parser/mode.
- Acceleration options: If you use speculative decoding (EAGLE3) and W4 quantization, you need to prepare the corresponding files and version requirements according to the instructions of the official repository/model card.
4. Typical use cases
- Serious consultation assistant: multiple rounds of questioning about symptoms, triggers, accompanying manifestations, past history and medication history, and output a structured summary and next step suggestions.
- Clinical auxiliary decision-making: Under the leadership of the doctor, give a list of differential diagnoses, recommended inspection items and risk warnings for "second opinions".
- Medical education and case discussion: Rewrite cases into standardized medical record points, and generate teaching questions and answers, key point reviews and knowledge point prompts.
- Medical content review: Check the consistency of popular science/consultation texts, and mark expressions that may not be rigorous or require evidence support.
5. Ecology and competing products
- Ecology: The basic model comes from Qwen3-235B-A22B, the training framework uses verl, and the inference side connects vLLM and SGLang, making it easy to fall into common open-source inference stacks.
- Competing products: Common routes to open source models for medical models include "continue pre-training + fine-tuning medical instructions" or "post-training based on validator/reward models". The difference between Baichuan-M3 is its emphasis on clinical process modeling and "fact-constrained RL". The evaluation set, data distribution, and compliance requirements of different organizations vary greatly, so it is recommended to do a comparative test within your real task and compliance boundaries.
6. Limitations and precautions
- It cannot replace professional diagnosis and treatment: The official clarifies that it is for research and reference only, and it is recommended to use it under the guidance of professional medical personnel.
- Evaluate extrapolated risks: Benchmark leadership does not mean that it is reliable for all departments/languages/populations, especially high-risk scenarios such as rare diseases, acute and critical illness, and drug dosage.
- High computing power and cost: The 235B scale has high requirements for video memory, bandwidth, and parallel strategy, and needs to be evaluated for latency, throughput, and cost before going online.
- Compliance and privacy: When it comes to medical records and personal information, data desensitization, access control, auditing, and human review processes are required.
7. Project address
https://huggingface.co/baichuan-inc/Baichuan-M3-235B
8. Frequently asked questions
Q: Is Baichuan-M3-235B really "less hallucinating and more diagnostic than GPT-5.2"?
A: The official comparison conclusion of HealthBench, HealthBench-Hard, hallucination evaluation and SCAN-bench is given in the model card; However, the evaluation settings and business distribution of different institutions vary greatly, so it is recommended to use your real case/consultation script for re-testing and manual review.
Q: Why did the Baichuan-M3-235B use Qwen3 as the base model?
A: The model is marked as Qwen3-235B-A22B in the model tree and acknowledgements, and its general capabilities such as large-scale MoE and long context are reused for medical backward training.
Q: What should I pay attention to when deploying Baichuan-M3-235B with vLLM?
A: Launch OpenAI-compatible services according to the official recommended version and enable qwen3's inference/parsing mode. The effects of multi-machine and multi-card parallelism, KV cache, context length and maximum output length on video memory are evaluated at the same time.
Q: How to choose between SGLang and vLLM deployment Baichuan-M3-235B?
A: Both are mainstream open-source reasoning frameworks; If you plan to use speculative decoding (such as Eagle3) or specific deployment parameters, you can first select the model according to the official example, and then compare the throughput, latency, and O&M complexity for stress testing.
Q: What role did verl play in the Baichuan-M3-235B training?
A: The official acknowledgment marks the training framework as verl; It is an open-source library for LLM post-training/RL and emphasizes integration with inference infrastructure such as vLLM, SGLang, and more.
