COLD START — Our endpoints scale to zero when idle to keep pricing at $0.001/request. If you're the first visitor in a while, the model warms up in 1–3 minutes. We show a live terminal feed with an ETA countdown and retry automatically — you never see a hard failure. Subsequent requests complete in seconds.
OMNIBIOFEX.CLOUD
DEMOCRATIZING WORLD-CLASS MEDICAL AI
Powered by
MEDGEMMA
1.5 4B IT
TRUST & COMPLIANCE
Every claim below is verifiable. Our platform inherits the certifications of its infrastructure partners — Google Cloud (hosting, database, inference), Razorpay (payments), and Resend (email). We publish exactly what we have, what we don't have, and what's still in progress. Honesty is a security feature.
55+Controls Verified
12Compliance Domains
256-bitEncryption Baseline
99.95%Uptime SLA
0PHI Retained
🔒
TLS 1.3 · 256-BIT
All traffic encrypted in transit with the latest TLS standard.
🛡️
AES-256 AT REST
Firestore + Cloud Storage encrypted with AES-256 by default.
🔐
CMEK READY
Customer-Managed Encryption Keys available for Enterprise.
🔑
GOOGLE OAUTH 2.0
Identity via Google. Zero passwords stored on our servers.
🎫
SHORT-LIVED JWT
Auto-rotated session tokens issued by Firebase Auth.
🚫
DEFAULT-DENY RULES
Firestore blocks all reads/writes unless explicitly permitted.
🛂
RBAC
Role-based access control — users read only their own data.
🔒
LEAST-PRIVILEGE
Every service, function, and user gets minimum required access.
🚫
ZERO PHI RETENTION
Uploaded files held in memory only for inference duration.
🧠
NO TRAINING ON DATA
Your uploads never train or fine-tune any model, ever.
🗑️
30-DAY DELETION
Full account and history deletion within 30 days of request.
🌍
MULTI-REGION REPLICATION
Firestore data replicated across regions for durability.
💳
PCI-DSS LEVEL 1
Highest level of payment card security certification.
💳
ISO 27001:2022
Razorpay holds current information security certification.
💳
SOC 2 TYPE 2
Independently audited security, availability, confidentiality.
💳
RBI-AUTHORISED PA
Reserve Bank of India Payment Aggregator license.
💳
AES-128 FIELD-LEVEL
Field-level encryption applied to sensitive payment data.
🚫
CARD DATA NEVER STORED
Card numbers never touch OmniBioFex servers — ever.
☁️
SOC 1 / 2 TYPE II / 3
Google Cloud holds all three SOC audit reports.
☁️
ISO 27001
Google Cloud information security management certified.
☁️
ISO 27017 / 27018 / 27701
Cloud-specific, PII, and privacy management certifications.
☁️
FedRAMP HIGH
Authorised at the highest federal impact level.
☁️
HIPAA-ELIGIBLE
BAA available via Google Cloud for PHI workloads.
⚡
99.95% UPTIME SLA
Published service level from Firebase + Cloud Functions.
🛡️
DDoS PROTECTION
Google Front End absorbs volumetric attacks at edge.
📧
SOC 2 TYPE II
Resend's transactional email infrastructure audited.
📧
EU-US DPF CERTIFIED
Data Privacy Framework compliant cross-border transfers.
📧
GDPR COMPLIANT
Email processing aligned with EU data protection law.
🇮🇳
DPDP ACT 2023 READY
India's Digital Personal Data Protection Act aligned.
🇪🇺
GDPR ARTICLE 9 ALIGNED
Special category health data handled with explicit consent.
🇺🇸
CCPA / CPRA READY
California consumer privacy rights supported.
🛡️
NIST CSF 2.0
Govern, Identify, Protect, Detect, Respond, Recover aligned.
🛡️
NIST SP 800-53
Control set inherited via Google Cloud infrastructure.
🛡️
OWASP ASVS
Application Security Verification Standard aligned.
🛡️
HITRUST CSF
Inheritable healthcare security controls via GCP.
🤖
ISO/IEC 42001
AI management system domains aligned (2023).
🤖
ISO/IEC 23894
AI risk management methodology aligned (2023).
🇪🇺
EU AI ACT ARTICLE 4
AI literacy obligations aligned for providers and deployers.
🌍
WHO AI ETHICS
Aligned with WHO guidance on AI ethics for health.
🌐
OECD AI PRINCIPLES
Inclusive growth, human rights, transparency aligned.
🤖
NIST AI RMF
AI Risk Management Framework mapped end-to-end.
🇺🇸
FDA GMLP
Good Machine Learning Practice guidance aligned.
🇸🇬
SINGAPORE AIHGLE 2.0
AI in Healthcare Guidelines (MOH/HSA) aligned.
📜
HAI-DEF GOVERNED
MedGemma under Google's Health AI Developer Foundations.
🏥
ISO 14971
Medical device risk management framework aligned.
🏥
IEC 62304
Medical device software lifecycle processes aligned.
🏥
ISO 13485
Medical device quality management system aligned.
🏥
ISO/TS 82304-2
Health software quality requirements aligned.
🔗
FHIR INTEROPERABILITY
HL7 FHIR readiness for EHR integrations.
🇬🇧
NHS DTAC / DSPT
Designed for UK NHS digital technology assessment.
🇬🇧
DCB0129 / DCB0160
UK clinical safety processes aligned.
🇮🇳
CDSCO SaMD PATHWAY
Aware of India's Medical Device Software regulatory path.
🛡️
CSA STAR LEVEL 2
Google Cloud third-party assessed by Cloud Security Alliance.
🇩🇪
BSI C5
German federal cloud security criteria — via Google Cloud.
🇫🇷
HDS CERTIFIED
French health data hosting certification — via Google Cloud.
🇸🇬
MTCS LEVEL 3
Singapore Multi-Tier Cloud Security — highest tier.
🇸🇬
OSPAR
Singapore financial services cloud audit — via Google Cloud.
🇪🇸
ENS CERTIFIED
Spain National Security Framework — via Google Cloud.
🇬🇧
CYBER ESSENTIALS PLUS
UK government-backed cybersecurity certification.
🇺🇸
DoD IL2 / IL4 / IL5
US DoD Impact Level provisional authorisations.
🇨🇭
ISAE 3000 TYPE 2
Swiss financial audit standard — via Google Cloud.
⚠️
NOT CLINICAL-GRADE
Outputs require independent verification before any clinical use.
⚠️
NOT FDA-CLEARED
Not a cleared medical device in the United States.
⚠️
NOT CE-MARKED
Not a marked medical device in the European Union.
⚠️
NOT HIPAA-COVERED
BAA pending. Do not upload identifiable PHI.
⚠️
RESEARCH USE ONLY
Educational and research use only. Not for clinical practice.
HOW TO READ THIS —
Infrastructure certifications (SOC, ISO, FedRAMP, HIPAA-eligible, CSA STAR, BSI C5, HDS, MTCS, OSPAR, ENS, Cyber Essentials, DoD IL, ISAE 3000) belong to Google Cloud, and OmniBioFex.Cloud inherits the controls but is not itself certified. Payment certifications belong to Razorpay. Email certifications belong to Resend. Frameworks marked "aligned", "ready", "aware of", or "designed for" indicate design intent and control mapping — not formal audit. We publish this list because the only trustworthy compliance claim is one you can verify.
WHAT WE PROVIDE
Everything you need to analyze, understand, and export medical scans with AI. From a single upload to a shareable report — no setup, no install, no credit card.
MULTIMODAL ANALYSIS
Upload X-rays, CT, MRI, histopathology slides, lab PDFs, or free-text EHR notes. One interface, every modality.
STRUCTURED REPORTS
MedGemma returns clean JSON with findings, peripheral margins, and impression — every output is a parseable artifact.
LIVE RESEARCH GROUNDING
Toggle once and every scan is cross-referenced against live PubMed, arXiv, and ClinicalTrials.gov — real, clickable citations.
CONVERSATIONAL FOLLOW-UP
After any analysis, ask follow-up questions about the report. MedGemma keeps your findings in context across the conversation.
VOICE INPUT
Dictate clinical context hands-free. Browser-native speech recognition converts your voice into the notes field instantly.
PHI DETECTION
Client-side scanner warns you before you send identifiable data. Detects SSN, Aadhaar, MRN, DOB, phone, and email patterns.
EXPORT & SHARE
Download any session as PDF, Markdown, or JSON. Follow-up chips and UI badges are stripped from exports — clean output.
DARK MODE + SHORTCUTS
Keyboard-first workflow (Ctrl+K, Ctrl+U, Ctrl+E) and a low-light theme built for radiologists reading in dark rooms.
AUTO-RETRY COLD START
Scale-to-zero infrastructure with automatic retry. If the GPU is cold, the platform retries up to 8 times — you never see a failure.
HOW IT WORKS
From sign-in to shareable report in under a minute. Google sign-in, drop your scan, get structured output — no configuration, no pipeline setup, no DevOps.
STEP 01
SIGN IN
One-click Google OAuth. No password to create, no email to verify, no credit card. Your account is live the moment you consent.
STEP 02
UPLOAD YOUR SCAN
Drag a chest X-ray, CT, MRI, or PDF. Multi-image uploads trigger longitudinal comparison. Crop the region of interest or use the full image.
STEP 03
AI ANALYZES
MedGemma 1.5 4B IT runs on Vertex AI. If Research mode is on, we fetch live PubMed, arXiv, and ClinicalTrials sources in parallel and inject them as context.
STEP 04
GET STRUCTURED OUTPUT
Receive a clean report: impression summary, peripheral margins analysis, discrete findings, and — if grounded — clickable citations to the exact papers used.
STEP 05
ASK, EXPORT, SHARE
Chat follow-ups in context. Download as PDF, Markdown, or JSON. Every session is stored in your history and retrievable anytime.
SPECIFICATION SHEET
MedGemma 1.5 4B IT is a developer starting point for healthcare AI applications. It provides strong baseline medical image and text comprehension for its size, but it is not clinical-grade and requires validation, adaptation, and/or fine-tuning before production use.
| Attribute | Specification |
|---|---|
| Base architecture | Gemma 3 decoder-only Transformer |
| Attention | Grouped-query attention (GQA) |
| Input modalities | Text, vision (multimodal) |
| Output modality | Text only |
| Context length | Supports long context, at least 128K tokens |
| Image input | Normalized to 896 × 896 resolution, encoded to 256 tokens each |
| Total input length | 128K tokens |
| Total output length | 8192 tokens |
| Training framework | JAX, optimized for TPUs |
| Image encoder | SigLIP, pre-trained on de-identified medical data |
ENCODER NOTE — The multimodal versions use a SigLIP image encoder specifically pre-trained on a variety of de-identified medical data, including chest X-rays, dermatology images, ophthalmology images, and histopathology slides.
CAPABILITIES OF 1.5
Expanding beyond ordinary medical QA and single 2D interpretation. MedGemma 1.5 4B IT supports new high-dimensional applications and complex reasoning tasks.
EHR UNDERSTANDING
Interpret text-based Electronic Health Record information. Allows conversion of longitudinal records into summaries, event extraction, and Q&A over patient records.
MEDICAL DOC → JSON
Extract structured information from unstructured medical laboratory reports (raw PDFs or images to JSON). A primary engineering use case.
3D MEDICAL IMAGING
Works natively with three-dimensional volume representations from CT and MRI scans. An essential evolution beyond flat 2D slice interpretation.
WHOLE-SLIDE HISTOPATHOLOGY
Simultaneously interprets multiple high-resolution patches taken from a whole-slide histopathology image (WSI).
LONGITUDINAL IMAGING
Interpret a current chest X-ray in the strict context of previous images, evaluating disease progression and temporal changes.
ANATOMICAL LOCALIZATION
Supports precise, bounding-box-based localization of medical features and anomalies directly within the text output.
DELTA VS MEDGEMMA 1 4B
IMPROVED MEDICAL TEXT REASONING
Higher accuracy on medical text reasoning compared with MedGemma 1 4B, alongside strengthened medical record interpretation and clinical reasoning baselines.
DELTA VS MEDGEMMA 1 4B
MODEST 2D IMAGE GAINS
Modest improvement on standard 2D image interpretation compared with MedGemma 1 4B. For image-only tasks — data-efficient classification, zero-shot classification, or content-based/semantic image retrieval — the MedSigLIP image encoder is recommended instead.
INTENDED USE CASES
MedGemma is an open multimodal generative AI model intended as a starting point for downstream healthcare applications involving medical text and images. Developers are responsible for training, adapting, and making meaningful changes for their specific use case. MedGemma can be fine-tuned with proprietary data.
4.1 — MEDICAL IMAGE INTERPRETATION
REPORT GENERATION
Generating medical image reports from radiology, pathology, dermatology, and ophthalmology inputs.
VISUAL Q&A
Answering natural language questions about medical images across multiple modalities.
MODALITY ADAPTATION
Adapting to CT, MRI, WSI, longitudinal imaging, and anatomical localization workflows.
4.2 — MEDICAL TEXT COMPREHENSION & CLINICAL REASONING
The model can be adapted for use cases requiring medical knowledge, such as patient interviewing, triaging, clinical decision support, summarization, medical document understanding and structured data extraction, and EHR interpretation. For most text-heavy use cases, the larger MedGemma 27B model generally yields the best performance. Developers should validate adapted models before deployment.
4.3 — ADAPTATION METHODS
PROMPT ENGINEERING
In-context learning: careful prompting, few-shot examples, or breaking tasks into subtasks.
FINE-TUNING
LoRA (parameter-efficient fine-tuning) and reinforcement learning. Fine-tune the language model decoder, the image encoder, or both.
AGENTIC ORCHESTRATION
Use MedGemma as a tool inside a larger system — coupled with web search, FHIR generators/interpreters, Gemini Live, or Gemini 2.5 Pro. It can parse private health data locally before sending anonymized requests to centralized models.
ACCESS PASSES
No subscriptions. No auto-renewal. Buy a pass when you need one. Skip a month if you don't.
COMMUNITY
$0/FREE
FOREVER FREEDesigned for medical students, educators, and evaluators.
- Quota: 10 analyses / month
- Model: Full MedGemma 1.5 4B access
- Modalities: X-ray, CT, MRI, histopath, EHR
- Research grounding: Not included
PRO · 30 DAYS
$29/PASS
30-DAY PASSUnlimited access. Live research grounding. Full feature set.
- Quota: Unlimited analyses
- Research: Live PubMed + arXiv + ClinicalTrials
- Extras: Voice, PHI detection, PDF/MD/JSON export
- Auto-Renew: None. Ever.
- (₹999 INR)
PRO · 90 DAYS
$75/PASS
SAVE $12Everything in 30-day. Three months. Better price.
- Effective: ≈ $25 / month
- Duration: 90 days from purchase
- Features: Same full Pro feature set
- Auto-Renew: None. Ever.
- (₹2,499 INR)
PRO · 365 DAYS ⭐
$249/PASS
BEST VALUE · SAVE $99The full year. Lowest per-month cost. Upfront one-time payment.
- Effective: ≈ $20.75 / month
- Duration: 365 days from purchase
- Features: Same full Pro feature set
- Auto-Renew: None. Ever.
- (₹7,999 INR)
API ACCESS · PREPAID WALLET
$0.001/REQ
Prepaid wallet · Minimum $10 · Never expires · Scale-to-zero · ₹0.095 per request
NO SUBSCRIPTIONS — Every Pro Pass is a one-time purchase. Your access runs for the exact duration shown. When it expires, your account automatically reverts to Community (10 analyses/month). No card is stored, no recurring charge is made, and you decide when to buy the next pass.
PERFORMANCE BENCHMARKS
MedGemma 1.5 4B was evaluated across multimodal classification, report generation, visual question answering, text-based tasks, medical record tasks, and document understanding — evaluated without safety filters.
5.1 — IMAGING EVALUATIONS
| Task / Dataset | Metric | MedGemma 1.5 4B |
|---|---|---|
| 3D Radiology | ||
| CT Dataset 1 (7 conditions) | Macro accuracy | 61.1 |
| CT-RATE (validation, 18 conditions) | Macro F1 | 27.0 |
| CT-RATE | Macro precision | 34.2 |
| CT-RATE | Macro recall | 42.0 |
| MRI Dataset 1 (10 conditions) | Macro accuracy | 64.7 |
| 2D Image Classification | ||
| MIMIC CXR | Macro F1 (top 5 conditions) | 89.5 |
| CheXpert CXR | Macro F1 (top 5 conditions) | 48.2 |
| CXR14 | Macro F1 (3 conditions) | 48.4 |
| PathMCQA (histopathology) | Accuracy | 70.0 |
| WSI-Path (whole-slide histopathology) | ROUGE | 49.4 |
| US-DermMCQA | Accuracy | 73.5 |
| EyePACS (fundus) | Accuracy | 76.8 |
| Disease Progression Classification (Longitudinal) | ||
| MS-CXR-T | Macro accuracy | 65.7 |
| Visual Question Answering | ||
| SLAKE (radiology) | Tokenized F1 | 59.7 |
| SLAKE | Accuracy (closed subset) | 82.8 |
| VQA-RAD (radiology) | Tokenized F1 | 48.1 |
| VQA-RAD | Accuracy (closed subset) | 70.2 |
| Region of Interest Detection | ||
| Chest ImaGenome: Anatomy bounding box detection | Intersection over union | 38.0 |
| Multimodal Medical Knowledge & Reasoning | ||
| MedXpertQA (text + multimodal questions) | Accuracy | 20.9 |
NOTE: MedGemma 1.5 4B shows strong radiology interpretation but was less optimized for the SLAKE Q&A format compared with MedGemma 1 4B. Fine-tuning on SLAKE may improve results.
5.2 — CHEST X-RAY REPORT GENERATION
| Task / Dataset | Metric | MedGemma 1.5 4B |
|---|---|---|
| MIMIC CXR – RadGraph F1 | RadGraph F1 | 27.2 |
For comparison, a fine-tuned MedGemma 1 4B achieved 30.3, and MedGemma 1 27B achieved 27.0. The instruction-tuned versions of MedGemma 4B and 27B achieve lower scores due to differences in reporting style compared with MIMIC ground-truth reports.
5.3 — TEXT EVALUATIONS
| Dataset | MedGemma 1.5 4B |
|---|---|
| MedQA (4-op) | 69.1 |
| MedMCQA | 59.8 |
| PubMedQA | 68.2 |
| MMLU Med | 69.6 |
| MedXpertQA (text only) | 16.4 |
| AfriMed-QA (25 question test set) | 56.0 |
5.4 — MEDICAL RECORD EVALUATIONS
| Dataset | Metric | MedGemma 1.5 4B |
|---|---|---|
| EHRQA | Accuracy | 89.6 |
| EHRNoteQA | Accuracy | 80.4 |
5.5 — DOCUMENT UNDERSTANDING EVALUATIONS
Evaluation of converting unstructured medical lab reports (PDFs/images) into structured JSON data.
| Task / Dataset | Metric | MedGemma 1.5 4B |
|---|---|---|
| EHR Dataset 2 (raw PDF to JSON) | Macro F1 | 91.0 |
| EHR Dataset 2 | Micro F1 | 88.0 |
| EHR Dataset 3 (raw PDF to JSON) | Macro F1 | 71.0 |
| EHR Dataset 3 | Micro F1 | 70.0 |
| Mendeley Clinical Laboratory Test Reports (PNG image of PDF to JSON) | Macro F1 | 85.0 |
| Mendeley Clinical Laboratory Test Reports | Micro F1 | 83.0 |
| EHR Dataset 4 | Macro F1 | 64.0 |
| EHR Dataset 4 | Micro F1 | 67.0 |
ETHICS & SAFETY EVALUATION
MedGemma 1.5 4B was evaluated for child safety, content safety, representational harms, and general medical harms.
CHILD SAFETY
Safe levels of performance observed compared with previous Gemma models.
CONTENT SAFETY
Minimal policy violations for both text-to-text and image-to-text tasks.
REPRESENTATIONAL HARMS
Safe levels observed across representational harm categories relative to earlier Gemma releases.
TESTING PROTOCOL — Testing was conducted without safety filters. The model produced minimal policy violations for both text-to-text and image-to-text tasks. LIMITATION: evaluations primarily used English-language prompts.
TRAINING DATA & DATA CARD
MedGemma 1.5 4B utilizes a combination of public and private datasets. The base Gemma models are pre-trained on a large corpus of text and code. The multimodal variants use a SigLIP image encoder pre-trained on de-identified medical data, including radiology, histopathology, ophthalmology, and dermatology images.
The LLM component is trained on diverse medical data, including medical text, medical question-answer pairs, FHIR-based EHR data (27B multimodal only), radiology images, histopathology patches, ophthalmology images, and dermatology images.
PUBLIC DATASETS
MIMIC-CXRChestImaGenomeSLAKEPAD-UFES-20SCINTCGACAMELYONPMC-OAMendeley Digital Knee X-RayMedQAMedMCQAPubMedQALiveQAAfriMed-QAVQA-RADMedExpQAMedXpertQAHealthSearchQAISICMendeley Clinical Laboratory Test ReportsCT-RATE
PRIVATE / PROPRIETARY DE-IDENTIFIED DATASETS
CT dataset 1MRI dataset 1Ophthalmology dataset 1 (EyePACS)Dermatology datasets 1–6Pathology datasets 1–4EHR datasets 1–5
For MedGemma 1.5 specifically, training and evaluation include expanded support for CT, MRI, WSI, longitudinal imaging, document understanding, and EHR data.
CRITICAL LIMITATIONS & USE
MedGemma 1.5 4B IT is NOT clinical-grade.
Outputs are not intended to directly inform clinical diagnosis, patient management, treatment recommendations, or direct clinical practice. All outputs should be considered preliminary and require independent verification, clinical correlation, and further investigation.
ADDITIONAL DOCUMENTED LIMITATIONS
MULTI-IMAGE GAPS
Multimodal capabilities were primarily evaluated on single-image tasks. Multiple-image comprehension has not been evaluated.
NO MULTI-TURN OPTIMIZATION
Not evaluated or optimized for multi-turn applications. May be more sensitive to specific prompts than Gemma 3.
BIAS & CONTAMINATION
Bias in validation data requires developers to validate on representative data. Data contamination is a risk — the model may have seen related medical information during pre-training, potentially overestimating generalization. Validate on datasets not publicly available where possible.
RELEASE NOTES & VERSIONS
MEDGEMMA 4B IT — CHANGELOG
MAY 20, 2025
Initial release.
JULY 9, 2025
Bug fix for subtle degradation in multimodal performance due to a missing end-of-image token in the model vocabulary. This impacted combined text-and-image tasks. The fix reinstates and correctly maps that token; text-only tasks remain unaffected.
JAN 13, 2026
Updated to version 1.5 with improved medical reasoning, medical records interpretation, and medical image interpretation.
JAN 23, 2026
Updated generation config to use greedy decoding by default. Sampling can still be allowed by users.
VERSIONS
| Resource ID | Release Date | Stage | Description |
|---|---|---|---|
| google/medgemma-1.5-4b-it | 2026-01-13 | GA | LLM for medical image and text comprehension with instruction tuning |
| google/medgemma-1.5-4b-it-dicom | 2026-01-13 | GA | Instruction-tuned LLM for medical image and text comprehension with DICOM support |
LICENSE & GOVERNANCE
MedGemma is governed by the Health AI Developer Foundations Terms of Use (last modified November 15, 2024) and the Health AI Developer Foundations Prohibited Use Policy.
KEY POINTS
- HAI-DEF recipe source code, model inference code, model training code, utility code, and other source code deliverables are licensed under Apache 2.0.
- Distribution of HAI-DEF or Model Derivatives requires inclusion of use restrictions and a copy of the Agreement.
- Users must not use HAI-DEF for prohibited uses, including illegal activities, harmful content, unauthorized professional practice, or generating deceptive content.
- Google will not claim ownership over original outputs generated using HAI-DEF.
- The services are provided "AS IS" without warranties. Google is not liable for damages arising from use.
- California law governs disputes, resolved in Santa Clara County courts.