Three-real training × long-horizon agents: the foundation we plan to build
Our technical plan: borrow general capability from open source and spend our resources on medical data, the training ground and validation. The planned 320B medical model would be adapted from an open-source base with three-real training rather than pre-trained from scratch; plan–act–verify–backtrack is meant to make 100-step clinical tasks dependable; every module will have a baseline and a validation method.
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
Models320B medical model (open-source base + medical continued pre-training) → 70B / 32B / 7B distillation
Data & computeProvincial data nodes (physically isolated) + federation and compute centre
What sets KunXi apart
Three-real training: an internship in a training ground
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 complete real tasks again and again inside replicas of hospital systems, and to score each decision against what actually happened to the patient afterwards.
Real data
authorised · longitudinal · multimodal
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
Key design choices
Training / validation / test split by patient, institution and time
Raw data and datasets stay in their domain; only model parameters are pooled
Seven modalities aligned on each patient’s timeline (see PTE)
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
Key design choices
One MCP tool layer for every business system
Expert demonstrations across 50 scenario types (planned)
Sandboxed, with actions graded A0–A4
Real outcomes
outcome-replay reinforcement learning
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
Key design choices
No future information: no peeking at the answer
Process rewards + rule verifiers + outcome rewards
A 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; demonstration volumes are planning figures
Technical route
A planned 320B medical model: to be adapted from an open-source base with three-real training, not pre-trained from scratch
NoticeThis module is a design concept, not yet built, and may change.
The route is a plan and may change: general capability would come from open source, with our resources going into medical data, the training ground and validation. Every module will have a baseline and a validation method.
Why adapt an open-source base
Pre-train 320B from scratchNot used
Cost and time
About 6.4 trillion tokens and 3.8 million GPU-equivalent hours (8× continued pre-training); ≥ 18 months
Dependencies and risks
Not enough medical data; heavy engineering load
Adapt an open-source baseChosen
Cost and time
Medical continued pre-training + supervised fine-tuning + reinforcement learning; two rounds in about 12 months (planned)
Dependencies and risks
Depends on open-source licences; the base is swappable, several candidates
External APIR&D only
Cost and time
Fastest to launch, pay per call
Dependencies and risks
Hospital data cannot leave; not controllable or auditable
Modules: what each solves, what it is compared with, how it is validated
← Swipe to see the full table →
Module
Problem it solves
Baseline
Validation
Stage
Retrieval-augmented generation
Rules, guidelines and drug labels cited live; every conclusion has a source
Model-only generation
Citation accuracy, factual error rate
Planned first release
Continued pre-training + fine-tuning
Hospital note language, coding habits and local rules
Open-source base, zero-shot
Coding agreement, edit volume
Planned first release
Agents + Harness
Task decomposition, tool use, step checks and approvals
Imaging, pathology, ECG and text reasoned together
Single-modality models
External test-set AUC
2028+
Engineering (planned)
70B on-premises (one 8-GPU server), 32B appliance, 7B offline; integrates HIS / EMR / LIS / PACS / insurance settlement lists; P95 ≤ 3 s per step (target); falls back to the rule engine on failure, staged release, one-click rollback.
SourcesChinchilla (2022); compute estimated as 6ND at 2.25 PFLOPS dense BF16 per GPU-equivalent and 40% utilisation; architecture and stages are KunXi plans; “planned first release” = modules the planned first product would need
Long-horizon tasks
End-to-end success on 100-step clinical tasks: from 13% to 90% (target)
NoticeThis module is a design concept, not yet built, and may change.
Overall success is per-step success to the 100th power. Plan–act–verify–backtrack is designed to raise the effective per-step success to 99.9% (target), jointly through the model and the Harness.
Task success vs number of steps (%)
98% per step
99% per step
98% per step + Harness catching 95% of errors
Why 13%
0.98¹⁰⁰ ≈ 13%; 0.99¹⁰⁰ ≈ 37%. Keep 98% per step but let the Harness catch 95% of the errors, and effective per-step success is about 99.9%, or about 90% over 100 steps.
Example: one inpatient stay ≈ 100 steps
Admission
Tests
Diagnosis
Treatment
Discharge summary
Coding & settlement
Follow-up
Planner · 320BBreaks down the task, orders the steps, re-plans on error
Process verifier · PRMScores every step; low scores backtrack or go to a person
Harness governanceA0–A4 levels, approvals and audit
step-by-step checks
78.2% vs 72.4%
Process reward models
Scoring every step beats scoring only the final answer
MATH benchmark
Lightman et al. 2023
15.6% → 77.9%
RL with verifiable rewards
Reward only answers that can be checked
Reasoning emerges on its own (AIME problems)
DeepSeek-R1 2025
67% in 400 steps
Tool-use reinforcement learning
Reasoning while calling a code tool
Text-only RL: 40% after 1,080 steps
ReTool · ByteDance Seed 2025
80%–85.5%
Multi-agent consultation
Diagnostic accuracy on 304 hard cases
General practitioners averaged 20%
MAI-DxO · Microsoft 2025
SourcesLightman et al. 2023 (process rewards); DeepSeek-R1 2025 (R1-Zero); ReTool (ByteDance Seed 2025); MAI-DxO (Microsoft 2025, SDBench 304 cases); these are public benchmark results; the 100-step figures are KunXi estimates and targets
Six frontier technologies
Standing on the global frontier, designing our own breakthroughs
NoticeThis module is a design concept, not yet built, and may change.
Each one follows the same structure: what the field has already shown, where the industry is stuck, how KunXi plans to design around it, which metric would validate it, and where it is meant to land. Methods, architectures, metrics and dates are KunXi design targets and concepts — not yet achieved, and subject to change.
Frontier technology 1Large models and agents
Rewards move from “right answer” to “real outcome”
NoticeThis module is a design concept, not yet built, and may change.
Reasoning models, reinforcement learning and agent orchestration already beat doctors in simulation; long-horizon reliability, authority to act and accountability in real clinical settings remain open.
Global frontierWHAT THE FIELD HAS SHOWN
85.5%
physicians averaged 20%
MAI-DxO sequential diagnosis orchestrator
304 hard NEJM cases; a virtual panel takes history, orders tests and diagnoses step by step under a cost budget
Microsoft AI · 2025-06
90%
final diagnosis in the top-7 differential
AMIE prospective real-clinic study
98 patients at Harvard-affiliated BIDMC; real-time human oversight, zero safety stops
Google · arXiv 2026-03
3.5%
hallucination rate
Baichuan-M3 fact-aware RL
Checks medical statements one by one; HealthBench 65.1 vs 63.3 for GPT-5.2-High
Baichuan · arXiv 2026-02
First
state-authorised autonomous AI refills
Utah AI regulatory sandbox
Doctronic renews routine prescriptions for chronic patients: AI gains a regulated “right to act”
Process rewards keep every step compliant; outcome rewards make the end result beneficial. Training runs in a closed loop inside the digital-twin ground and enters clinical workflow only after graded Harness release.
1Digital-twin training groundReplicas of HIS / EMR / LIS / insurance; replays years of real cases inside the authorised domain
2Policy model320B planning + 70B execution; multi-step tool calls, each step yielding an auditable intermediate conclusion
3Process reward (PRM)Step-by-step checks against guidelines, 15,500 insurance rules and drug labels
5Graded Harness releaseA0–A4 action levels; conformal abstention; full audit logs
↻Policy update (GRPO): process and outcome rewards flow back into the policy model in a closed loop
Industry bottleneckPlanned KunXi approachValidation metric (target)
Rewards depend on exam answers and expert preference, detached from patient outcomes
Build verifiable rewards from years of longitudinal real outcomes
≥ 85%MedAgentBench task success (2027)
Errors compound over long tasks; 100-step success below 15%
Step-by-step process rewards + backtracking and re-planning on failure
≥ 90%100-step task completion (2028)
No grading, approval or accountability for AI actions
A0–A4 action levels, approval chain and full audit
0S3 critical errors (monitored continuously)
Where it is meant to landThe planned first product, insurance compliance & record QC (70B on-premises) → six kinds of digital workers
SourcesMicrosoft AI, “The Path to Medical Superintelligence” (2025-06-30); arXiv 2603.08448 (AMIE prospective study, 2026-03); arXiv 2602.06570 (Baichuan-M3, 2026-02); Utah Department of Commerce (2026-01-06); Med-PRM (EMNLP 2025); Microsoft Research arXiv 2609.12277; methods, architectures and metrics are KunXi design targets
Frontier technology 2Multimodality and precision medicine
Pathology and genomics have foundation models; next comes whole-course alignment
NoticeThis module is a design concept, not yet built, and may change.
Single-modality, single-time-point models are maturing and entering the clinic. Data that spans modalities and time and is paired with years of outcomes is scarce, and that is what holds precision medicine back from “accurate diagnosis” to “precise treatment”.
Global frontierWHAT THE FIELD HAS SHOWN
2.3M
whole-slide pathology images
PRISM2 multimodal pathology model
Trained on 200,000 patients and 14 million Q&A pairs; pan-cancer detection AUC 0.967, prompt-driven reasoning
Nature Medicine · 2026-07
77%
vs 61% for PD-L1 expression
MUSK vision-language oncology model
Predicts immunotherapy benefit in NSCLC; pre-trained on 50 million pathology images + 1 billion text tokens
Nature · 2025-01
25 / 26
variant-effect benchmarks matched or led
AlphaGenome genome model
1 million bases per input; predicts 5,930 human genome function signals at once
Nature · 2026-01
78,600
real-world opportunistic screenings
GRAPE gastric-cancer screening on plain CT
External validation across 16 centres, AUC 0.927; sensitivity 21.8% higher than radiologists
Nature Medicine · 2025-06
KunXi design conceptWHAT WE PLAN TO BUILD
PTE · patient timeline encoder
Concept · planned from 2027
For each patient, imaging, pathology, labs, notes, medication and follow-up are aligned in time and anchored to years of real outcomes to learn a single representation that supports reasoning about treatment response and recurrence risk.
Where it is meant to landPrecision-oncology RWE and companion population screening for pharma; MDT agents for hospitals; to be delivered jointly with encoder partners
SourcesNature Medicine 32:3203 (PRISM2, 2026-07-31); Nature 638 (MUSK, 2025-01-08); Nature 649:1206 (AlphaGenome, 2026-01-28); Nature Medicine 31:3011 (GRAPE, 2025-06-24); Ruijin RuiPath 2.0 in use at 90+ hospitals in China (Xinhua, 2026-08-31); methods, architectures and metrics are KunXi design targets
Frontier technology 3Prediction and prevention
From predicting risk to simulating interventions: counterfactual trajectories for personal prevention
NoticeThis module is a design concept, not yet built, and may change.
Generative trajectory models already predict over a thousand diseases at once and show scaling laws. But a risk score does not say which intervention to choose, and large-scale screening still has to prove benefit on hard endpoints.
Global frontierWHAT THE FIELD HAS SHOWN
1,258
diseases predicted at once
Delphi-2M health-trajectory model
Trained on 400,000 UK Biobank participants, externally validated on 1.9 million Danes; horizon up to 20 years
Nature · 2025-09
118M
patients’ longitudinal records
Epic CoMET medical-event model
Pre-trained on 115 billion medical events; zero-shot it matches or beats specialised models on most of 78 tasks
Epic · arXiv 2025-08
−60%
calibration error (ECE 0.082→0.033)
Reinforcement learning on patient trajectories
After GRPO / DAPO fine-tuning, a 9M model’s AUROC beats a 56M pre-trained model
Microsoft Research · 2026-09
−14%
stage IV cancer diagnoses
NHS-Galleri multi-cancer screening RCT
142,000 people; fewer stage IV cases but the primary endpoint was missed — predictions must be proven on hard endpoints
GRAIL · ASCO 2026
KunXi design conceptWHAT WE PLAN TO BUILD
CPT · counterfactual patient twin
Concept · validation planned 2028
On an event-trajectory model, candidate interventions are generated conditionally and the resulting outcome distributions compared; target-trial emulation corrects for confounding and real outcomes keep it calibrated.
1Event-trajectory pre-trainingVisits, labs, medication and outcomes encoded as events; years of population-wide regional records
Calibration drifts after deployment; risk is over- or under-estimated
Outcome-replay RL + monthly recalibration and drift monitoring
≥ 80%Directional agreement with published RCTs
Where it is meant to landPricing and underwriting for insured people with conditions; personal interventions in the chronic-care companion; risk stratification for medical alliances
SourcesNature (Delphi-2M, 2025-09-17); arXiv 2508.12104 (Epic Cosmos generative medical-event model); arXiv 2609.12277 (Microsoft Research, 2026-09); GRAIL 2026-05-30 (NHS-Galleri, ASCO 2026); risk values in the figure are illustrations, not predictions; methods, architectures and metrics are KunXi design targets
Frontier technology 4Pharma R&D and real-world evidence
The regulatory framework is in place; KunXi plans a re-runnable evidence factory
NoticeThis module is a design concept, not yet built, and may change.
The FDA asks for AI credibility evidence graded by context of use and model risk; trial emulation on EHR data already agrees with randomised trials. The bottlenecks are cycle time, reproducibility and pooling data across institutions.
Global frontierWHAT THE FIELD HAS SHOWN
7 steps
credibility assessment framework
FDA AI credibility guidance (draft)
Graded by context of use (COU) and model risk; requires a credibility assessment plan and report
FDA · 2025-01
9 / 9
comparable trials statistically equivalent
TRIALSCOPE real-world trial emulation
RWD from 1 million+ cancer patients emulates 11 NSCLC trials, over 20× faster
Microsoft · Providence · 2025
−42.6%
screening time for enrolment
TrialGPT trial matching
An LLM checks each inclusion and exclusion criterion; 87.3% accuracy, close to experts
Agents run the whole chain: protocol parsing, federated cohorts, causal inference, bias audit and the evidence pack. Code, data versions and decision logs are recorded throughout, so any conclusion can be re-run with one click.
1Protocol-parsing agentExtracts PICO, time zero and endpoints from the protocol (RAG + human check)
2Federated cohort buildGenerates phenotype code automatically; queries and counts in parallel on provincial nodes
Negative controls, E-values and sensitivity analyses by default
≥ 80%directional agreement when reproducing past RCTs
Data is spread across institutions and raw data cannot be pooled
Federated execution on provincial nodes; statistics only
−40%screening time for enrolment (pilot target)
Where it is meant to landEvidence packs, trial-recruitment assistants and external controls for pharma; to be delivered with CROs
SourcesFDA Draft Guidance (2025-01); arXiv 2311.01301 (TRIALSCOPE); Nature Communications (TrialGPT, 2024-11-18); Nature Medicine (rentosertib, 2025-06); EMA PROCOVA qualification (2022-09); medRxiv 2026-04 (LLM target-trial emulation, 92.3% accuracy extracting protocol elements); methods, architectures and metrics are KunXi design targets
Frontier technology 5Trusted computing and data as an asset
Provincial isolation + a federation centre: data stays in the province, value still flows
NoticeThis module is a design concept, not yet built, and may change.
China is building trusted data spaces, and the question for medical data has moved from “whether” to “how to run it”. Confidential computing, differential privacy and federated learning make cross-province collaboration workable.
Global frontierWHAT THE FIELD HAS SHOWN
100+
trusted data spaces (by 2028)
National trusted data space action plan
July 2026: a national programme on medical data use, now focused on roll-out and operations
National Data Administration · NDRC
< 15%
records accessed across hospitals
Medical data that “can’t leave the hospital”
About 70% is unstructured; proposals at the 2026 Two Sessions urge medical trusted data spaces
Two Sessions · 2026-03
< 7%
confidential-computing inference overhead
GPU confidential computing (TEE), measured
Most LLM queries add under 7%; 70B models are almost unaffected; the bottleneck is CPU–GPU transfer
arXiv 2409.03992
ε ≤ 2
sequence-level differential privacy
VaultGemma differentially private model
A 1-billion-parameter open model trained from scratch with DP; no memorisation of training data detected
Google Research · 2025-09
KunXi design conceptWHAT WE PLAN TO BUILD
FedHub · medical data federation
Concept · first nodes planned 2027
Provincial data nodes are physically isolated and raw data never leaves the province; federated computing and learning run locally, while the federation centre handles secure aggregation, 320B training, audit records and value sharing.
Provincial node 1
Provincial node 2
Provincial node 3
Federation + compute centreSecure aggregation · TEEAudit records320B training · value sharing
Provincial node 4
Provincial nodes 5–7
Provincial node N
Raw data physically isolated · stays in the provinceEncrypted parameters ↑ global model ↓
Industry bottleneckPlanned KunXi approachValidation metric (target)
Raw data cannot be pooled across institutions or provinces
Physically isolated provincial nodes exchange encrypted parameters only
0 bytesof raw data leaving a province
Parties don’t trust each other and computation is hard to prove
TEE confidential computing + remote attestation + differential privacy
≤ 3%performance gap, federated vs centralised training
Data value cannot be measured or shared
Contribution metering + smart-contract settlement + blockchain records
Where it is meant to landFederated training services, data products and trusted-data-space operations
SourcesNational Data Administration, Action Plan for Trusted Data Spaces (2024–2028) (2024-11-23); NDRC (2026-07-08); Shanghai Securities News (2026-03-09); arXiv 2409.03992; Google Research (VaultGemma, 2025-09-12); arXiv 2601.22124 (Fed-MedLoRA); architecture and metrics are KunXi design targets
Frontier technology 6Digital health and omics
Continuous signals are new biomarkers; KunXi plans to join life outside and inside the hospital
NoticeThis module is a design concept, not yet built, and may change.
Foundation models for wearables and continuous glucose show that long-term continuous signals predict disease. The gap: signals from outside the hospital are cut off from hospital records and outcome labels.
Global frontierWHAT THE FIELD HAS SHOWN
2.5B hours
of wearable behaviour data
Apple wearable behaviour model (WBM)
162,000 participants, 57 health tasks; fused with PPG, it detects pregnancy with 92% accuracy
Apple · arXiv 2025-07
66%
of new diabetes over 12 years caught
GluFormer continuous-glucose model
Trained on 10 million+ CGM readings; the top-risk quartile also caught 69% of cardiovascular deaths
Nature · 2026-01
92.3%
specificity (sensitivity 41.2%)
Apple Watch hypertension notifications
Passive optical screening cleared by the FDA: consumer devices enter regulated medicine
FDA · 2025-09
600,000
plasma proteomics samples
UK Biobank proteomics project
Up to 5,400 proteins per sample; a handful predict 67 diseases better than routine clinical measures
UK Biobank · 2025-01
KunXi design conceptWHAT WE PLAN TO BUILD
DCF · dual-clock fusion of life outside and inside the hospital
Concept · launch planned 2027
Minute-level continuous signals from outside the hospital and month-level hospital records are aligned on each person’s timeline; hospital diagnoses and follow-up outcomes label the outside signals, giving a personal baseline and early warning.
1Continuous signals outsideWatches / bands, CGM, home blood-pressure monitors, sleep and activity
27B on-device modelLocal pre-processing and anomaly detection; raw signals stay on the device
4320B personal reasoningPersonal baseline and risk tiers; A0 record · A1 reminder · A2 referral suggestion
5Closing the loopDoctor confirms, return visits and follow-up, insurer wellness incentives
Omics (proteomics, genomics) is tested by partners; KunXi fuses it with the clinical timeline for high-risk review and precision cohorts for pharma
Industry bottleneckPlanned KunXi approachValidation metric (target)
Outside signals and hospital records are disconnected, with no outcome labels
Dual-clock alignment labelled by hospital diagnoses and outcomes
+10 ppBP / glucose control rate (prospective validation)
Consumer sensors are imprecise (41.2% sensitivity for hypertension alerts)
Trend tiers and reminders only; positives confirmed in hospital
≥ 5device brands connected (2027)
Omics tests are costly and cover few people
Partner integration for precision cohorts and high-risk review
Where it is meant to landThe To C chronic-care companion; partnerships with wearable makers; insurer wellness incentives
SourcesarXiv 2507.00191 (Apple WBM, 2025-07); Nature (GluFormer, 2026-01); FDA / Fierce Biotech (Apple hypertension notifications, 2025-09-17); UK Biobank (2025-01-10); Nature Medicine 2025 (PH-LLM); methods, architectures and metrics are KunXi design targets
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.
Build it with us
Explore putting a medical LLM to work on real clinical tasks
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.
