---
title: "Baichuan Bet the Company on Medical AI — and Its Model Claims a 3.5% Hallucination Rate"
date: 2026-11-22
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-baichuan-medical-hallucination
language: en
---

# Baichuan Bet the Company on Medical AI — and Its Model Claims a 3.5% Hallucination Rate

> Wang Xiaochuan abandoned general models for healthcare; the test is whether low-hallucination claims survive contact with real clinics.

Ask a general chatbot about a rare disease and it may produce a fluent answer with a fabricated citation. In consumer software that is annoying. In a clinic it is dangerous. One Chinese model company decided that reducing this error — not widening general knowledge — was the product.

Baichuan Intelligent Technology (百川智能) was founded in 2023 by Wang Xiaochuan (王小川), previously the founder of search firm Sogou. In April 2025, after two years in the crowded general-model race, Wang narrowed the company to healthcare. In January 2026 it open-sourced Baichuan-M3, a medical model, and in May it released Baichuan-M4 alongside an AI family-doctor product called Bai Xiaoying (百小应).

## Key takeaways

- Baichuan pivoted from general large models (大模型) to medical AI in April 2025.

- Baichuan-M3 was open-sourced on 13 January 2026; M4 followed in May with a family-doctor application.

- The company reported a medical hallucination rate of 3.5% for M3 at launch, using its own evaluation.

- The training method, called fact-aware reinforcement learning, builds factual consistency into the reward signal.

- Benchmark leadership claimed by the company has not been confirmed by independent clinical trials.

## The specific problem

Medical questions are unusually hostile to language models. The answer must be current, sourced, and careful about uncertainty. A plausible but wrong drug interaction is worse than no answer, and patients cannot tell the difference.

Baichuan's approach, as the company describes it, is to make factual consistency part of reinforcement learning rather than a filtering step added afterwards. The idea is that a model rewarded for factual alignment will learn to hedge when evidence is thin, instead of filling gaps with confident text. The company also reports using retrieval and evidence-anchoring so that conclusions can be traced to journal articles or guideline passages.

This is a different philosophy from general assistants, which are optimised for helpfulness and breadth. A medical model that says "I don't know" more often is, by the company's logic, a better product.

## Reading the benchmark claims carefully

At the M3 launch in January, Baichuan said the model ranked first on HealthBench, a medical benchmark from OpenAI, and on the harder HealthBench Hard subset. It also reported a medical hallucination rate of 3.5%, and said M3 was available through its Bai Xiaoying (百小应) assistant.

These are company-reported results. Benchmarks measure performance on fixed question sets, which never fully represent clinical practice. A model can top HealthBench and still fail when a patient describes symptoms in messy, incomplete language. The honest interpretation is that Baichuan has published an encouraging internal result, not that clinical superiority has been established.

The same caution applies to the M4 announcement. Media coverage of the May launch describes a collaboration with Tsinghua University and reports low hallucination figures across several benchmarks. Again, those numbers originate with the developer.

## The commercial bet

Wang has said the company may start an IPO process in 2027 and that it held roughly ¥3 billion (≈ US$420M) in cash at the time of the January interview. That cushion matters because medical AI monetises slowly. Hospitals buy through procurement cycles, require regulatory clearance, and expect evidence from real clinical use.

The product strategy reflects patience. Rather than issuing diagnoses or prescriptions, which Chinese regulation restricts, the tools help patients understand treatment options and prepare for appointments. Wang has described an initial free consumer product, with revenue later coming from service packages or partnerships with drug and device companies. That is a plausible route, but it depends on reaching a large user base before charging.

The pivot has a cost. Baichuan was once grouped with several other well-funded Chinese model startups. Competitors that stayed in general models have listed or raised at higher valuations. Baichuan accepted a narrower position in exchange for a domain where it believes it can be defensible.

## Why this matters beyond one company

Medical AI is where the abstract debate about hallucination becomes concrete. If models are to be used in health systems, the industry needs shared standards for measuring factual error, not each vendor's preferred benchmark. A company publishing its numbers, even self-reported, adds pressure for standardisation.

There is also a language dimension. Clinical language models trained mostly on English transfer imperfectly. Chinese medical terminology, drug names, and practice patterns require local data. A domestic model built for Chinese hospitals may outperform a globally famous one on the tasks that matter locally — but that must be demonstrated, not assumed.

## Honest limitations

- All benchmark rankings and hallucination rates in this article are company-reported. I found no independent clinical validation published in a peer-reviewed journal.

- Cash reserves and IPO timing come from a single interview reported by Jiemian News, not an audited disclosure.

- The M4 launch details are drawn from Chinese media descriptions of the event. I did not obtain the technical report.

- I have excluded several widely circulated figures from self-media articles because they could not be traced to primary sources.

- HealthBench measures a defined task set. High scores do not establish safety in real clinical use.

## What readers can do now

- If you use an AI tool for health questions, check whether it cites sources you can open and verify. A confident answer without a traceable source should not change a treatment decision.

- Treat "ranked first" as a statement about a benchmark, not about your doctor's problem. Ask which benchmark and who ran it.

- For practitioners, insist on local validation. A model evaluated elsewhere may not reflect your patient population or prescribing rules.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-baichuan-medical-hallucination
Free to quote with attribution and a link to the original.
