Product preview · KunXi Medical LLM · by Great Unicorn Intelligence
KunXi Medical LLM is a line of large models and agents for healthcare that Great Unicorn Intelligence is designing. The plan: a 320B medical model acts as the teacher, and three-real training (real data, real environments, real outcomes) lets agents practise 100-step clinical tasks inside replicas of hospital systems. The goal is to return a finished task with a traceable evidence chain, not just a paragraph of text.
Product previewKunXi Medical LLM is at the design and concept stage
This section describes what we plan to build and how we intend to do it — not finished products or results. Product forms, model sizes, technical approaches, features, interfaces, metrics, scenarios and timelines are design concepts that have not been built or validated; they may be adjusted or changed as development, validation and regulatory requirements evolve. Nothing here is a commitment, an offer or a claim of clinical effectiveness.
All rights reserved. No reposting, adaptation or commercial use without written permission.Read the copyright notice
Whole-course timeline · seven modalities
5 years agoToday
Inpatient journey task
Evidence chainevery conclusion → source note · guideline · 15,500 insurance rules
Conformal abstentionlow confidence → “I’m not sure” → a human takes over
HarnessA3 action awaiting insurance-office approval · full audit
NoticeThis module is a design concept, not yet built, and may change.
Swap a case vignette for a multi-turn consultation and diagnostic accuracy drops sharply; asked to work inside an EHR, the best model succeeds only about 70% of the time. The root cause: training is detached from the clinic.
CRAFT-MD: GPT-4, 2,000 cases, 12 specialties
Models are trained on question banks, textbooks and single-hospital data: static, single-modality, with no real outcomes. KunXi plans to address this with three-real training.
SourcesCRAFT-MD (Nat Med 2025); MedAgentBench (NEJM AI 2025); HealthBench Hard (OpenAI 2025; Baichuan-M2 2025); CREOLA (npj Digit Med 2025); NMPA registrations (ChinaMedGlobal, secondary source)
NoticeThis module is a design concept, not yet built, and may change.
On the left, what published studies show today; on the right, the technical answer KunXi plans to use (a design concept, not yet built). Targets are design metrics and must be validated one by one before release.
01Long-horizon tasks
Best EHR task success is about 70%; at 98% per step, 100 steps succeed only 13% of the time
Hierarchical planning + process reward model + Harness checks and backtracking; target ≥ 90% over 100 steps
02Multimodality
Approved AI software is mostly single-disease, single-modality; lung and cardiovascular make up 43%
A whole-course timeline: seven modalities encoded together and reasoned over by one model
03Real consultations
In multi-turn consultations, diagnostic accuracy falls from 82% to 63%
Patient simulators + reinforcement learning on clinical rubrics
04Hallucination & safety
1.47% of sentences in AI-written notes are hallucinated; 44% of those are major
Evidence tracing + conformal abstention + A0–A4 graded approval
05Hard, long-tail cases
On a physician-graded hard set, the best score is still below 0.5
Multi-agent consultation + specialty routing + active learning on the long tail
06Data silos
Data sits across hospitals, primary care, insurance and public health; raw data cannot leave its domain
Multi-domain federated learning + differential privacy: the data stays, the model moves
NoticeThis module is a design concept, not yet built, and may change.
By design, KunXi is not meant to be a general Q&A bot. The plan is to put a medical LLM to work in the real operations of hospitals, pharma companies, insurers and families: reading records, checking rules, verifying orders, drafting summaries, scheduling follow-up. In the design, every step traces back to source evidence, and the closer an action gets to a clinical decision, the stricter its approval.
What sets KunXi apart
NoticeThis module is a design concept, not yet built, and may change.
The usual approach teaches a model to pass exams. We plan to have it do an “internship” inside replicas of hospital systems, then score every decision against what actually happened to the patient afterwards.
Usual approachQuestion banks, textbooks, papers, single-hospital data
KunXi (planned)Planned: population-wide longitudinal regional data, used only within lawful authorisation: hospitals, primary care, insurance and public health joined up, years of history, seven modalities; raw data never leaves its domain
7 modalitiesyears of history · patient-level timelines
Usual approachSingle-turn Q&A, static exam questions
KunXi (planned)Planned: replicas of HIS, LIS, PACS, insurance and pharmacy systems inside the authorised domain, where agents would practise 100-step tasks in a sandbox again and again
100 stepsper task · expert demonstrations across 50 scenario types (planned)
Usual approachMatching reference answers or human preferences
KunXi (planned)By design, the model only sees data up to a given moment and decides; it is then scored on what really followed: readmissions, complications, targets met and cost
Years of follow-upa time moat: outcome data takes years of follow-up to build
| Today’s medical LLMs | KunXi: three-real training (planned) | |
|---|---|---|
| What it learns from | Question banks, textbooks, papers | Authorised population-wide longitudinal multimodal data |
| How it practises | Single-turn Q&A, static questions | 100-step hands-on tasks in a digital-twin training ground |
| Who scores it | Reference answers, human preference | Real outcomes + rule verifiers + process rewards |
| Target output | A paragraph of text | Goal: a finished task + a traceable evidence chain |
SourcesMedAgentBench (NEJM AI 2025) evaluates agents in a virtual EHR, the academic precedent for a “training ground”
NoticeThis module is a design concept, not yet built, and may change.
In the plan, the 320B medical model plans long tasks and acts as the teacher; it would be distilled into 70B, 32B and 7B tiers for the cloud, hospitals, appliances and offline devices in primary care. Model sizes and tiers may change.
Long-task planning, RWE evidence packs, claim re-review; teacher for the whole family
Insurance offices, medical-records departments and wards of tertiary hospitals; the planned first product
Secondary hospitals, medical alliances, check-up centres, pharmacy-chain HQs
Primary care, pharmacy stores, wearables and consumer devices
NoticeThis module is a design concept, not yet built, and may change.
The four sub-sections answer four questions: what we plan to deliver, how we intend to get there, where it is meant to be used, and how we will test whether it works.
Planned: six kinds of To B digital workers + a To C chronic-care companion; the planned first product is an insurance-compliance and record-QC agent; a planned 320B / 70B / 32B / 7B model family.
Insurance complianceInpatient journeyEvidence engineerUnderwriting & claimsPrescription reviewHealth managementChronic careOpen
Three-real training, a 320B route built on an open-source base, 100-step long-horizon tasks and KunXi design concepts for six frontier technologies.
RLVOPTECPTREFFedHubDCFOpen
The plan: start with in-hospital insurance compliance and record QC, then extend to hospitals, pharma, insurers, pharmacies, health management and families.
HospitalsPharmaInsurersPharmaciesHealth managementFamiliesOpen
Designed five-stage validation; evidence tracing, conformal abstention, A0–A4 graded actions; raw data and datasets stay in their domain.
Five stagesS0–S3 error gradesA0–A4Egress listOpenNoticeThis module is a design concept, not yet built, and may change.
We plan to validate technical ability, workflow improvement, clinical benefit, safety and economic value separately; exam scores are a floor for capability, never evidence of clinical effectiveness.
Retrospective data; compared with general models, rule engines and people under the same conditions
Another hospital, another period; patients, institutions and time kept apart
Three months in the background with no clinical effect; every case compared with the human result
“Doctor + AI” versus “doctor”: time, edits, misses
Change in denials and deductions; diagnostic aids follow device clinical-trial rules
Note: KunXi Medical LLM is a product preview at the design and concept stage. Unless a public study is cited, the products, capabilities, thresholds, metrics and timelines in this section are KunXi design targets, concepts or estimates — not achieved or measured results — and may change; product screens are concept illustrations with simulated data. By design, the agents only make suggestions; clinical decisions are made by doctors.
KunXi Medical LLM is at the design and concept stage. Hospitals, pharma companies, insurers, pharmacies and health-management providers who are interested are welcome to talk to us and assess data, workflow and the validation path together.
