A patient types "I have a headache." Most chatbots would spew a paragraph about brain tumors and end with "please see a doctor." Baichuan Intelligence (百川智能) built something that does the opposite: it pauses, asks which side hurts, how long it has lasted, whether there is fever or blurred vision — and only then reasons.
Now picture the second scene. A rural clinic has no neurologist on duty at night. The family describes symptoms into a phone. The model does not guess a diagnosis; it walks through the same structured questions a triage nurse would, flags the red flags, and tells the family whether this is "monitor at home" or "go to the emergency room now." That gap — between a confident answer and a careful conversation — is exactly where Baichuan-M3 (百川-M3) is designed to live.
The model is an open-source medical large model (大模型) released by the Beijing-based company on 13 January 2026. It is not a general-purpose chatbot wearing a medical skin; it is trained, from the ground up, to behave like a clinician taking a history.
Key takeaways
- Baichuan-M3 is an open-source medical large model (大模型) released on 13 January 2026 by Baichuan Intelligence (百川智能).
- On OpenAI's HealthBench it scored 65.1 overall and 44.4 on the harder HealthBench Hard subset, figures the company says lead the public ranking.
- The company reports a 3.5% hallucination rate — the lowest it claims among current medical models.
- The model's defining feature is native end-to-end (端到端) consultation: it asks follow-up questions, narrows the differential, and reasons over complete information rather than answering in one shot.
- It is already wired into Baichuan's "Baixiaoying" (百小应) app for both doctors and patients.
Why "asking back" matters in medicine
Hallucination is annoying in a chatbot and dangerous in a clinic. A wrong restaurant recommendation costs a diner an evening; a wrong medication suggestion can cost a life. Baichuan's bet is that the safest way to cut medical hallucination is not to bolt on a retrieval tool at the end, but to push factual consistency into training itself. The company describes a reinforcement-learning setup where medical fact consistency is a core training objective, so the model learns "to know what it does not know" as part of its own behavior.
The result the company highlights is a 3.5% hallucination rate in its own pure-model setting — no external retrieval, no tool calls. Baichuan frames this as a deliberate architectural choice: keep the model honest by training it to be honest, rather than patching honesty on afterward. The contrast with a typical search box is sharp. Search returns plausible text; a careful clinician returns a process. M3 is optimized to reproduce the process.
The consultation loop
Where most medical assistants answer immediately, Baichuan-M3 is built to run a loop. It proposes a SCAN principle the company describes as safety layering, information clarification, associated follow-up, and normalized output. In practice: the patient gives a vague symptom, the model asks targeted questions, extracts key history and risk signals, then performs deeper reasoning on the fuller picture.
The company describes this as a philosophy it associates with founder Wang Xiaochuan (王小川): the goal is to shift decision power from doctors to patients, so people can understand their own diagnoses and treatment plans. Whether that cultural shift arrives is a separate question; the engineering point is that the model is optimized for longitudinal, multi-turn medical dialogue rather than single-turn Q&A.
The capability is already in production. Baichuan's medical application "Baixiaoying" (百小应) has integrated M3, offering both a doctor mode and a patient mode so clinicians can rehearse reasoning and families can make sense of diagnoses. In a country where specialists are concentrated in tier-one cities, a tool that asks the right questions at the edges of the health system is where the real value sits.
The wider Chinese medical-AI race
Baichuan is not alone in the medical-AI field. What separates Baichuan is focus. The company spent 2025 narrowing from a broad model portfolio to medicine specifically — chronic disease, pediatrics, and oncology are the stated priorities. That focus is why a single medical model, rather than a general flagship, became its headline release.
The open-source decision matters for adoption. Hospitals and clinics in China are cautious about sending patient data to closed external APIs. A model they can host internally, inspect, and fine-tune on local data lowers the barrier to real deployment — provided the license and clinical-use boundaries are clear.
What is actually verifiable
The scores are real numbers published by the company and echoed by Chinese state-affiliated English outlets. HealthBench is OpenAI's medical evaluation suite; HealthBench Hard stresses complex decision-making. Baichuan says M3 leads both. On the low-hallucination claim, the 3.5% figure is the company's own measurement in a pure-model configuration.
The company also claims M3 surpasses GPT-5.2 on these benchmarks — a claim I report as Baichuan's, not as an independently confirmed fact, because I have not re-run the evaluations.
What readers can do now
- If you build or buy medical AI, test multi-turn behavior, not single-turn answers. A model that asks one good follow-up question is far safer than one that answers instantly and confidently.
- Treat "open-source" as a starting point, not a certificate. Baichuan-M3 is open-weight; verify the license and the intended clinical-use boundaries before any deployment near patients.
- Watch the hallucination metric, but ask how it was measured. A 3.5% rate in a company's internal pure-model test is not the same as a 3.5% rate in your hospital's data.
Honest limitations
This article relies on figures published by Baichuan Intelligence and reported by Chinese state-affiliated English media (China Youth Network's English service) and Guangzhou Daily's online outlet. The HealthBench 65.1 / HealthBench Hard 44.4 scores and the 3.5% hallucination rate are company-stated and were not independently re-measured here. The claim that M3 surpasses GPT-5.2 is presented as the company's assertion. I have not reviewed the arXiv preprint (arXiv:2602.06570) in detail, and peer-reviewed clinical validation is outside the scope of this column. I have not compared M3 against other Chinese medical models such as those from Dingxiang (丁香) or Huawei-backed efforts, which would be needed for a full market picture.