---
title: "Unique Partner | Tsinghua Team Yizhe AI Builds Health-Management AI Agents with MoE LLMs"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2023-11-02T17:00:00+08:00"
canonical: "https://ffcap.cn/en/research/20231102a0798c00"
source: "https://ffcap.cn/en/research/20231102a0798c00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20231102A0798C00"
---

# Unique Partner | Tsinghua Team Yizhe AI Builds Health-Management AI Agents with MoE LLMs

_Unique Research · Insight / Unique Partner welcome · November 2, 2023 · KEEP partner welcome; inventory type insight · Unique Capital as FA on new raise_

AI + healthcare is a broad growth market. Global Market Insights: global medical AI ~**$5B** in 2022 → **$70B** by 2032, CAGR >29%. IQVIA: China medical AI only ~**RMB 4B** in 2021, but next decade CAGR >30%. The trend will mint new AI+healthcare leaders.

## Tsinghua team building health-management AI agents

**Yizhe AI (医者AI)** is rising in the field. Founded **2023** by Tsinghua CS PhDs, MDs, and successful AI-health founders. CEO **Liu Chenghui** is a serial AI+health entrepreneur; CTO **Xu Zhenan** is a Tsinghua CS PhD deep in medical NLP, founder of Tsinghua Student Algorithm Association, co-author of China Electronics Society technical group standards, international awards, deployments in top tertiary hospitals. Yizhe AI is raising a new round with **Unique Capital as financial advisor**.

Focus: sub-health management. Latest **MoE**\-architecture LLMs plus hardware/software deliver 7×24 health management — every family its own health-steward AI and Healthy Care Agents.

Market solutions today split into human services vs. traditional tech products — both flawed. Human service: high-level clinicians concentrate in hospital systems, ignoring long-tail health management. Traditional expert systems / knowledge graphs are keyword-triggered Q&A bots — weak on complex daily health scenes and unable to build trust.

## Self-built Lego-style modular MoE LLM

Foundation models can cut cost and offer humanlike high-level interaction — but plain fine-tuned or industry models still hit hallucination, weak long-term memory, and thin few-shot training. Yizhe’s answer: self-built MoE.

MoE fuses multiple expert nets; a gating model routes tasks to expert “squads,” spawning new AI agents. New needs do not require full retrain — low marginal cost, tiny samples → new agent models.

Yizhe’s MoE: ~**70B** total params, **10+** experts (diabetes, “three highs,” pulmonary nodules, etc.). Examples: general health AI routes “inquiry” + functional experts; nodule AI routes compound judgment + treatment + nodule experts; diabetes nutritionist AI routes diabetes + dietetics experts.

Four differences vs. fine-tuned LLMs: (1) consistency/accuracy >**97%**, cutting hallucination; (2) strong long-term memory — up to **3 years** of dialogue history; (3) few-shot — multilayer architecture; ~**1,000** few-shot rounds can clone a specialist model; (4) proactive guidance via 1+9 MoE — complex-step replay, dialogues lasting >**20 minutes**.

## Scenes

**Checkups:** partnered with leading private checkup groups — pre-checkup “AI online family doctor” plans; post-checkup report interpreter AI + specialist AI; personal health digital twin for 7×24 monitoring. ~**12M** user volume in this scene.

**Family doctor:** with leading insurers — WeChat-group online doctor model; dedicated AI family doctor plus coach/nutritionist agent team. Radiating **500k+** families.

Also multi-end (phone, PC, smart speaker, watch) for a product company’s post-diabetes management scene — 7×24 care anywhere.

Building user mindshare and general model apps toward efficient health management for tens of millions of Chinese families. Plan: DTC in **Q3 2024**; **1M** C-end users by **Q2 2025**.

_1 CN image; attach on ingest. KEEP partner-welcome tone; inventory type **insight**._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20231102a0798c00
On-site reading page: https://ffcap.cn/en/research/20231102a0798c00