Zhizhen launches WiseDiagV3 and HaoBan AI, pushing medical AI from single answers to continuous care
新智元 · wechat · 2026-07-28
WiseDiagV3 and Good Companion AI shift medical AI from single-turn answers to continuous service
Zhizhen Technology says its new WiseDiagV3 medical model and upgraded HaoBan AI move medical AI beyond “one-off replies” into a continuous health service model.
What changed
- Better medical reasoning: V3 expands continued pretraining data to 100B tokens, covering papers, guidelines, medical Q&A, clinical data, textbooks, and more, with supervision from 50 medical experts.
- Multi-specialty fusion: the post-training stack adds OPD expert fusion and online distillation so the model can synthesize views across specialties.
- Generative reward model: a self-built GRM scores responses on accuracy, completeness, safety, and clarity to reduce hallucinations.
Memory and family health management
- HaoBan AI can now identify family members in natural language.
- It stores structured health facts such as history, meds, allergies, and test results.
- It uses confirmation cards so users can approve, delay, or revoke writes to long-term memory.
- It can also create a new member profile automatically when a new family member is mentioned.
Agentic workflow
The new Z-loop service flow: identify the problem, ask targeted follow-ups, pull relevant context, coordinate multi-agent reasoning, and return an actionable report.
Example use case
The article says the system helped reason through a complicated case that ultimately pointed to pulmonary arteriovenous malformation (PAVM) after low oxygen and a stroke with no obvious cardiac source. The suggested next steps were CT angiography, possible embolization, and coordinated anticoagulation review.
The company frames this as a broader shift in medical AI evaluation: from “how strong is the model” to “how complete is the service.”
More from Models
- Kimi K3 goes live on Modal as an open 3T-class model — songhan_mit · 2026-07-28
- Inference.net pitches a gateway flow that mirrors prod traffic to Kimi K3 before switching — MatthewBerman · 2026-07-28
- Claude was unsubscribed as ChatGPT/Codex 5.6, Sol and Kimi 3 all struggled — sull · 2026-07-28
- GLM 5.5 is said to arrive in August with stronger long-horizon agent loops — bindureddy · 2026-07-28
- Moonshot’s Kimi K3 lands in Japan with 2.8T open weights and $3/$13 pricing — DavidBennett__ · 2026-07-28
- Users report DeepSeek’s web search is degrading and mixing languages on every query — teortaxesTex · 2026-07-28