Diagnify-14B
Medical reasoning connected to our curated evidence layer.
v0.2-CoT · access-gated · in testing
Medical models · clinician decision support
Medical AI.
Built for clinical reasoning.
Our clinically fine-tuned models combine medical reasoning with source-linked evidence. Three connected products support clinicians, from patient history to signed note.
Research model Diagnify-14B v0.2-CoT Fine-tuned for medicine Open-weight base · clinical adaptation Evidence at every step FINE-TUNE → RETRIEVE → GROUND → CHECK Designed for clinician review
Clinician-reviewed
Evidence-grounded
In testing
Clinical reasoning, in one workspace.
01
What we build
Three products. One workflow.
Take the history, work up the case, and sign the record — together or as separate tools.
Clinician console
Clinical decision support
Evidence-linked differentials, investigations and plans that update as findings arrive.
- Red flags first
- Differential updates live
- Sources beside recommendations
Diagnify Intake
Voice history before the appointment, in the patient’s language.
- Secure link; no app
- Twelve languages
- Structured history for review
Diagnify EMR
Intake, scribe and decision support inside your practice record.
- Your practice subdomain
- Notes, referrals and scripts
- Clinician signs the record
02
Our models
Medical models. Clinical tasks.
We fine-tune medical models and ground their reasoning in a clinician-curated evidence base. The programme spans language, ECGs, skin images and examination findings.
Diagnify ECG
Reconstructs 12-lead signals from images for structured interpretation.
Clinician-reviewed second read
Dermatology Vision
Skin-image recognition for structured examination findings.
In training · not yet released
Examination Vision
Normal-versus-abnormal visual findings for clinician confirmation.
In training · not yet released
Outputs can be wrong. Clinician review is required.
03
Results
A 14B model. 99.9% in our internal evaluation.
Tested on 2,500+ complex cases, graded blind by five senior physicians. The best comparator reached 83%.
Internal retrospective evaluation; not peer-reviewed. Chart percentages are rounded. GPT-4o is a separate 2025 reference. Full comparison and limits are below.
+17 pts
above the best comparator
GPT-5.5 and Claude Opus 4.8: 83%.
~2,500
clinician-confirmed cases
De-identified cases; blinded physician grading.
0%
safety-critical misses in this cohort
Comparator range: 8–19%.
04
Head-to-head results
Tested against five AI models.
Same cases. Blinded grading. Confidence intervals shown.
99.8–100% (95% CI), internal evaluation
+17 pts over the next-highest model
p<0.001 vs all five comparators
No safety-critical miss observed in this cohort
Graded by five senior physicians
| Diagnostic accuracy · 2,500+ cases | Correct | 95% CI | vs DIAGNIFY |
|---|---|---|---|
| DIAGNIFY Medical LLM structured expert mode | 99.9%#1 | 99.8–100% | — |
| GPT-5.5 OpenAI | 83% | 81.5–84.5% | p<0.001 |
| Claude Opus 4.8 Anthropic | 83% | 81.5–84.5% | p<0.001 |
| Gemini 3.1 Pro Google DeepMind | 75% | 73.3–76.7% | p<0.001 |
| Grok 4.20 xAI | 58% | 56.1–59.9% | p<0.001 |
| DeepSeek V4 DeepSeek | 42% | 40.1–43.9% | p<0.001 |
| GPT-4o historical reference · 2025 study, not re-tested | 49% | — | — |
Unpublished internal, retrospective evaluation run by DIAGNIFY; not peer-reviewed or externally replicated. Results do not establish clinical effectiveness or patient outcomes. Five comparators used public APIs with default settings. GPT-4o is a separate 2025 reference, not re-tested.
Study methods and subgroup results
Where the accuracy gap was biggest — rare and multi-system cases.
Pre-specified subgroups: ultra-rare, multi-system, safety-critical misses, and investigation efficiency. The margin over comparators was larger in these subgroups than in the primary outcome.
Ultra-rare diseases
450 cases · prevalence <1 in 50,000
100%
vs 28–71% across the field
Multi-system disease
880 cases · ≥3 organ systems involved
100%
vs 78–79% best comparator
Safety-critical misses
life-threatening dx left off the differential
0%
vs 8–19% across the field
Unnecessary tests
investigations ordered without discriminating value
3%
vs 12–31% across the field
How the study was run.
Blinded grading, paired testing on an identical cohort, and cases post-dating the models.
5
senior physicians, blinded
Graders across internal medicine, clinical genetics, neurology, oncology and vascular medicine; inter-rater agreement Cohen's κ = 0.94 (0.92–0.96).
p<0.001
vs each comparator
McNemar's test for paired proportions with Bonferroni correction for multiple comparisons; significant against all five models.
2024–26
post-dates the models
Cases were published after the models were built and confirmed by histopathology, genetics, imaging or specialist consensus; de-identified in accordance with the Privacy Act 1988 (Cth) and the Australian Privacy Principles.
05
Specialty coverage
Built specialty by specialty.
Green: in testing. Orange: on the roadmap.
In testing
Coming soon
General Practice
Emergency Medicine
Behavioural Paediatrics
General & Internal Medicine
Cardiology
Respiratory & Sleep Medicine
Gastroenterology & Hepatology
Endocrinology & Diabetes
Nephrology
Neurology
Rheumatology
Haematology
Medical Oncology
Infectious Diseases
Clinical Immunology & Allergy
Geriatric Medicine
Palliative Medicine
General Paediatrics
Neonatology
Psychiatry
Dermatology
Intensive Care Medicine
Anaesthesia
Pain Medicine
Rehabilitation Medicine
Sport & Exercise Medicine
Addiction Medicine
Sexual Health Medicine
Occupational & Environmental Medicine
Public Health Medicine
Clinical Genetics
Nuclear Medicine
Diagnostic Radiology
Clinical Pharmacology
06
Decision support
Three ways to work a case.
Type, talk or use the consultation transcript. The same reasoning engine powers each.
Interactive
Work through a case as new findings update the differential and plan.
Voice
Discuss the case, challenge the differential and review management hands-free.
Ambient scribe
Turn the consultation transcript into structured findings and decision support.
07
How it works
Reasoning you can follow.
Red flags, differential, investigations and evidence checks — on one reviewable path.
STEP 01 · SCREEN
Screen
Raise red flags and must-not-miss conditions first.
STEP 02 · REASON
Reason
Update the differential as findings arrive.
STEP 03 · WORK UP
Work up
Choose investigations that distinguish the leading possibilities.
STEP 04 · VERIFY
Check
Check sources, doses and contraindications before clinician review.
08
Application
Everyday and chronic care.
In testing for non-urgent consultations and chronic-disease reviews, including rural and remote care.
History
Structured history-taking
Capture red flags and the questions that distinguish likely diagnoses.
Self-examination
Guided self-examination
Guide defined home checks and record findings for clinician review.
Visual examination
Single-frame visual examination
Use a single camera frame at the examination step.
Reasoning
Same reasoning architecture
Review the differential, work-up and assessment together.
Scope
In testing. Clinician oversight required.
09
Before the appointment
History ready before the appointment.
Patients answer an adaptive history in their language. Clinicians receive a structured summary.
→
History only
No diagnosis, triage category or treatment advice.
→
Clinician confirms
Review and correct the history during the consultation.
→
Emergency guidance
Anyone acutely unwell is directed to call 000.
10
Privacy & data residency
Your patients' data stays in Australia.
✓
Australian GPU compute
GPU inference runs in Sydney.
✓
Sydney data residency
Patient data is stored in Sydney and stays in Australia.
✓
Encrypted records
Encrypted in transit and at rest.
Important — please read
Decision support, in testing. A registered practitioner must verify every output. Benchmark results are not clinical validation or regulatory approval. In an emergency, call 000.