GitHub - maziyarpanahi/openmed: Local-first healthcare AI: clinical NER & HIPAA PII de-identification that runs 100% on-device. 2,200+ medical models, 21 languages, Apple MLX + Python, no cloud, no patient data leaving your network. Apache-2.0

Your Data. Your Model. Your Hardware.

Turn clinical text into structured, de-identified insight, with nothing uploaded.
OpenMed extracts biomedical entities and removes 55+ PHI types entirely on the hardware you control, so your data never leaves the device. The same 2,000+ open models run from a phone to a GPU server, fully offline: iOS, iPadOS, and Android via OpenMedKit, React Native, plain CPUs, Apple Silicon, NVIDIA GPUs, the browser, and REST/gRPC services. No cloud. No vendor lock-in. No patient data leaving your network.

2,000+ models · 24 model-backed PII languages · 600+ PII checkpoints · 100% on-device · Apache-2.0

English · 简体中文 · Español · Français · Deutsch · Italiano · Português · Nederlands · العربية · हिन्दी · తెలుగు · 日本語 · Türkçe · فارسی · Kiswahili

See it in action

OpenMed runs entirely on the device; clinical text never leaves it. Here it is on iPhone, fully offline:

_{On iPhone via OpenMedKit: scan a clinical note, de-identify it, and extract clinical signals, all locally with Apple MLX. Nothing is uploaded.}

_{Real-time PII de-identification: the Nemotron Privacy Filter redacting names, addresses, IDs, and billing data from a clinical discharge packet, entirely on-device. (All values shown are synthetic.)}

30-second example

from openmed import analyze_text

result = analyze_text(
    "Patient started on imatinib for chronic myeloid leukemia.",
    model_name="disease_detection_superclinical",
)

for entity in result.entities:
    print(f"{entity.label:<12} {entity.text:<28} {entity.confidence:.2f}")
# DISEASE      chronic myeloid leukemia     0.98
# DRUG         imatinib                     0.95

A state-of-the-art clinical NER model running locally: no API key, no network call.

Why OpenMed?

	OpenMed	Cloud medical APIs
Runs on your device / servers	✅	❌
Patient data leaves your network	Never	Sent to the vendor
Cost	Free & open-source	Per-call pricing
Specialized medical models	2,000+	Limited
Model-backed PII languages	21	Varies
Offline / air-gapped	✅	❌
Apple Silicon (MLX) acceleration	✅	n/a
Native iOS / macOS apps	✅ via OpenMedKit	❌
Browser/WebGPU token classification	✅ via Transformers.js	Varies
Vendor lock-in	None (Apache-2.0)	Yes

Specialized models: 2,000+ curated biomedical & clinical models, many outperforming proprietary stacks.
HIPAA-aware de-identification: all 18 Safe Harbor identifiers, smart entity merging, format-preserving fakes.
Runs everywhere: CPU, CUDA, Apple Silicon (MLX), iOS/macOS via OpenMedKit, Android/Kotlin, React Native, REST/gRPC services, and browser/WebGPU bundles via Transformers.js.
One-line deployment: Python API, Dockerized REST service, or batch pipelines.
Zero lock-in: Apache-2.0, your infrastructure, your data.

On-device on Apple: Swift, MLX & iOS

OpenMed is built to run where your data already lives. On Apple hardware it accelerates with MLX, and it ships straight into iPhone, iPad, and Mac apps through OpenMedKit: so PII detection and clinical extraction happen fully offline, on the device.

// Add OpenMedKit to your app
dependencies: [
    .package(url: "https://github.com/maziyarpanahi/openmed.git", from: "1.9.1"),
]

Expected result: Swift Package Manager resolves OpenMedKit and makes import OpenMedKit available to your app target.

MLX runtime for PII token classification, the Privacy Filter family, experimental GLiNER-family zero-shot tasks, and Python MLX-LM text generation with Laneformer; includes a CoreML fallback path for supported token-classification artifacts.
One model name, every platform: MLX model names automatically fall back to the matching PyTorch checkpoint on non-Apple hardware.
Python on Apple Silicon too: pip install --upgrade "openmed[mlx]".

Guides: MLX backend · OpenMedKit (Swift) · CoreML export

_{MLX on Apple Silicon: 24–33× faster than CPU PyTorch for the Privacy Filter: median latency per inference step, lower is better.}

On-device on Android — Kotlin & ONNX Runtime Mobile

OpenMedKit also ships as a native Android/Kotlin library for local document intake, OCR handoff, PII redaction, and token-classification inference through ONNX Runtime Mobile. Mobile model repositories include stable tensor names, dynamic sequence axes, tokenizer files, labels, and Android-ready fp32, fp16, INT8, and optional .ort outputs.

Add the scoped JitPack repository in settings.gradle.kts:

dependencyResolutionManagement {
    repositories {
        google()
        mavenCentral()
        maven {
            url = uri("https://jitpack.io")
            content { includeGroup("com.github.maziyarpanahi") }
        }
    }
}

Then use the immutable OpenMed v1.9.1 release:

dependencies {
    implementation("com.github.maziyarpanahi:openmed:v1.9.1")
}

See the Android installation guide for local builds and publishing details.

val model = OpenMedKit.fromDirectory(modelDir)
val entities = model.analyzeText("Patient Alice Nguyen was seen in cardiology.")

Android ONNX profile emits model.onnx, model_fp16.onnx, model_int8.onnx, tokenizer assets, labels, and openmed-onnx.json.
ORT Mobile support records the minimal-build operator configuration when ONNX Runtime conversion tooling is installed.
Kotlin parity tests keep tokenizer offsets, span boundaries, and decoder output aligned with the Python runtime.

Guides: Android ONNX export · Android span parity · OpenMedKit Android

The same ONNX model on Python CPU

from openmed import OnnxModel

model = OnnxModel.from_pretrained(
    "OpenMed/OpenMed-PII-ClinicalE5-Small-33M-v1-onnx-android"
)
entities = model("Patient Alice Nguyen was seen in cardiology.")

The same ONNX model in the browser

npm install openmed @huggingface/transformers

import { loadOnnxModel } from "openmed";

const model = await loadOnnxModel(
  "OpenMed/OpenMed-PII-ClinicalE5-Small-33M-v1-onnx-android",
);
const entities = await model("Patient Alice Nguyen was seen in cardiology.");

How it works

flowchart LR
    A["Clinical text"] --> B["OpenMed<br/>(100% on-device)"]
    B --> C["Medical entities"]
    B --> D["PII detected"]
    B --> E["De-identified text"]
    style B fill:#0D6E6E,stroke:#0A5656,stroke-width:2px,color:#ffffff
    style C fill:#D6EBEB,stroke:#0D6E6E,color:#0E1116
    style D fill:#F7DCD8,stroke:#C5453A,color:#0E1116
    style E fill:#F5E27A,stroke:#A9A088,color:#0E1116

Rendered result: a local clinical-text pipeline that returns medical entities, PII findings, and de-identified text without sending data to a cloud API.

Quick start

# Core + Hugging Face runtime (Linux, macOS, Windows; CPU or CUDA)
pip install --upgrade "openmed[hf]"

# Add the REST service
pip install --upgrade "openmed[hf,service]"

# Apple Silicon acceleration (MLX)
pip install --upgrade "openmed[mlx]"

Expected result:

Successfully installed openmed-...

Python API

from openmed import analyze_text

result = analyze_text(
  "Patient received 75mg "
  "clopidogrel for NSTEMI.",
  model_name=
  "pharma_detection_superclinical",
)
print([(e.label, e.text) for e in result.entities])

Example output:

[('DRUG', 'clopidogrel'), ('CONDITION', 'NSTEMI')]

REST service

uvicorn openmed.service.app:app \
  --host 0.0.0.0 --port 8080

Example output:

INFO:     Uvicorn running on http://0.0.0.0:8080
GET /health -> 200 OK

GET /health POST /analyze POST /pii/extract POST /pii/deidentify

Batch

from openmed import BatchProcessor

p = BatchProcessor(
  model_name=
  "disease_detection_superclinical",
  group_entities=True,
)
results = p.process_texts([...])
print(len(results), sum(len(r.entities) for r in results))
print([(e.label, e.text) for e in results[0].entities[:1]])

Example output:

3 7
[('DISEASE', 'leukemia')]

Browser / WebGPU

Package ONNX token-classification exports for in-browser inference through Transformers.js:

python -m openmed.onnx.convert \
  --model dslim/bert-base-NER \
  --output dist/example-onnx \
  --include-transformersjs

Example output:

Exported Transformers.js bundle to dist/example-onnx

import { pipeline } from "@huggingface/transformers";

const detector = await pipeline(
  "token-classification",
  "/models/openmed-pii/transformersjs",
  { device: "webgpu" },
);
const entities = await detector("Patient Casey Example called 212-555-0198.");
console.log(entities.slice(0, 2));

Example output:

[
  { entity: "NAME", word: "Casey Example", score: 0.99 },
  { entity: "PHONE", word: "212-555-0198", score: 0.98 },
]

Transformers.js export guide

Offline / air-gapped? Point model_name (or model_id) at a local directory and OpenMed loads it without contacting the Hugging Face Hub:

from openmed import OpenMedConfig, analyze_text

result = analyze_text(
    "Patient presents with chronic myeloid leukemia and Type 2 diabetes.",
    model_id="./models/OpenMed-NER-DiseaseDetect-SuperClinical-434M",
    config=OpenMedConfig(device="cpu"),
)
for entity in result.entities:
    print(f"{entity.label:<12} {entity.text:<28} {entity.confidence:.2f}")

Example output:

DISEASE      chronic myeloid leukemia     0.98
DISEASE      Type 2 diabetes              0.96

Because model_id points to a local directory, this example does not contact the Hugging Face Hub or any external model provider.

Models

A curated registry of specialized medical NER models; browse the full catalog.

Model	Specialization	Entity types	Size
`disease_detection_superclinical`	Disease & conditions	DISEASE, CONDITION, DIAGNOSIS	434M
`pharma_detection_superclinical`	Drugs & medications	DRUG, MEDICATION, TREATMENT	434M
`pii_superclinical_large`	PII & de-identification	NAME, DATE, SSN, PHONE, EMAIL, ADDRESS	434M
`anatomy_detection_electramed`	Anatomy & body parts	ANATOMY, ORGAN, BODY_PART	109M
`gene_detection_genecorpus`	Genes & proteins	GENE, PROTEIN	109M

Privacy: PII detection & de-identification

from openmed import extract_pii, deidentify

text = "Patient: John Doe, DOB: 01/15/1970, SSN: 123-45-6789"

# Extract PII with smart merging (prevents tokenization fragmentation)
result = extract_pii(text, model_name="pii_superclinical_large", use_smart_merging=True)
print([(e.label, e.text) for e in result.entities])

# De-identify with the method you need
print(deidentify(text, method="mask").deidentified_text)
print(deidentify(text, method="replace").deidentified_text)
print(deidentify(text, method="hash").deidentified_text)
print(deidentify(text, method="shift_dates", date_shift_days=180).deidentified_text)

Example output:

[('NAME', 'John Doe'), ('DATE', '01/15/1970'), ('SSN', '123-45-6789')]
Patient: [NAME], DOB: [DATE], SSN: [SSN]
Patient: Emily Chen, DOB: 03/22/1985, SSN: 456-78-9012
Patient: 6b8f...c4a1, DOB: 48b1...91de, SSN: 3f13...e912
Patient: John Doe, DOB: 07/14/1970, SSN: 123-45-6789

Smart entity merging keeps 01/15/1970 whole instead of fragmenting it.
Policy-aware pipelines add HIPAA/GDPR/research profiles, calibrated thresholds, signed audit reports, redaction previews, and minimum-necessary action selection.
Faker-backed obfuscation with custom clinical-ID providers (CPF, CNPJ, BSN, NIR, Codice Fiscale, NIE, Aadhaar, Steuer-ID, NPI).
HIPAA: all 18 Safe Harbor identifiers, configurable confidence thresholds.
Batch and streaming PII: extract or de-identify across many documents with BatchProcessor(operation="extract_pii" | "deidentify", batch_size=16) or incremental streaming helpers.

Batch PII processing throughput: up to 3.3x on CPU and 2.2x on MLX

_{Batch processing: up to 3.3× higher throughput on CPU and 2.2× on MLX vs. one document at a time.}

Complete PII notebook · Smart merging · Anonymization quickstart

Privacy Filter family: three model families on the OpenAI Privacy Filter architecture

Same model code (gpt-oss-style sparse-MoE transformer with local attention, sink tokens, RoPE+YaRN, tiktoken o200k_base), different training data. All route through the same extract_pii() / deidentify() API; only model_name= changes. openai/privacy-filter is a Hugging Face model identifier for local weights; using it here does not call the OpenAI API.

Variant	PyTorch (CPU + CUDA)	MLX (Apple Silicon)	MLX 8-bit
OpenAI Privacy Filter	`openai/privacy-filter`	`OpenMed/privacy-filter-mlx`	`…-mlx-8bit`
Nemotron-PII fine-tune	`OpenMed/privacy-filter-nemotron`	`…-nemotron-mlx`	`…-nemotron-mlx-8bit`
OpenMed Multilingual	`OpenMed/privacy-filter-multilingual`	`…-multilingual-mlx`	`…-multilingual-mlx-8bit`

from openmed import extract_pii

text = "Patient Sarah Connor (DOB: 03/15/1985) at MRN 4471882."

variants = {
    "baseline": extract_pii(text, model_name="openai/privacy-filter"),
    "nemotron": extract_pii(text, model_name="OpenMed/privacy-filter-nemotron"),
    "mlx": extract_pii(text, model_name="OpenMed/privacy-filter-mlx"),
}
print([(e.label, e.text) for e in variants["baseline"].entities])

Example output:

[('NAME', 'Sarah Connor'), ('DATE', '03/15/1985'), ('ID', '4471882')]

On non-Apple-Silicon hosts, MLX model names are automatically substituted with the matching PyTorch checkpoint (with a one-time warning): ship one model name, run anywhere. See Privacy Filter architecture & backend routing.

Multilingual PII (26 supported languages)

Extraction and de-identification support 26 supported PII language codes: am, ar, da, de, en, es, fr, he, hi, id, it, ja, ko, nl, no, pt, ro, ru, sv, sw, te, th, tr, xh, zh, and zu, with 600+ PII checkpoints in total. Russian and Chinese routing currently use documented multilingual default-model placeholders while dedicated model weights remain separate. An optional, user-configured Indic NER family accepts nine additional routes (as, bn, gu, kn, ml, mr, or, pa, and ta) and can also serve Hindi and Telugu. Set OPENMED_INDIC_NER_MODEL; OpenMed never bundles or automatically selects those weights. OpenMed also includes validator-backed national-ID coverage for additional ID-only locales such as Polish, Latvian, Slovak, Malay, Filipino, and Finnish.

See the per-language guide for each code's default PII model, Faker locale, and a before/after de-identification example.

python -c "from openmed import extract_pii; print([(e.label, e.text) for e in extract_pii('Dr. Pedro Almeida, CPF: 123.456.789-09, email: pedro@hospital.pt', lang='pt').entities])"

Example output:

[('NAME', 'Pedro Almeida'), ('ID', '123.456.789-09'), ('EMAIL', 'pedro@hospital.pt')]

Show per-language examples (Portuguese, Dutch, Hindi, Arabic, Japanese, Turkish)

from openmed import extract_pii

portuguese = extract_pii("Paciente: Pedro Almeida, CPF: 123.456.789-09, telefone: +351 912 345 678", lang="pt", use_smart_merging=True)
dutch      = extract_pii("Patiënt: Eva de Vries, BSN: 123456782, telefoon: +31 6 12345678", lang="nl", use_smart_merging=True)
hindi      = extract_pii("रोगी: अनीता शर्मा, फोन: +91 9876543210, पता: नई दिल्ली 110001", lang="hi", use_smart_merging=True)
arabic     = extract_pii("المريضة ليلى حسن، الهاتف +20 10 1234 5678، الرقم القومي 29801011234567.", lang="ar", use_smart_merging=True)
japanese   = extract_pii("患者 佐藤 花子、電話 +81 90 1234 5678、マイナンバー 1234 5678 9012.", lang="ja", use_smart_merging=True)
turkish    = extract_pii("Hasta Ayşe Yılmaz, telefon +90 532 123 45 67, TCKN 10000000146.", lang="tr", use_smart_merging=True)

for r in (portuguese, dutch, hindi, arabic, japanese, turkish):
    print([(e.label, e.text) for e in r.entities])

Example output:

[('NAME', 'Pedro Almeida'), ('ID', '123.456.789-09'), ('PHONE', '+351 912 345 678')]
[('NAME', 'Eva de Vries'), ('ID', '123456782'), ('PHONE', '+31 6 12345678')]
[('NAME', 'अनीता शर्मा'), ('PHONE', '+91 9876543210'), ('ADDRESS', 'नई दिल्ली 110001')]
[('NAME', 'ليلى حسن'), ('PHONE', '+20 10 1234 5678'), ('ID', '29801011234567')]
[('NAME', '佐藤 花子'), ('PHONE', '+81 90 1234 5678'), ('ID', '1234 5678 9012')]
[('NAME', 'Ayşe Yılmaz'), ('PHONE', '+90 532 123 45 67'), ('ID', '10000000146')]

REST API

A Docker-friendly FastAPI service with request validation, shared pipeline preload, and unified error envelopes.

pip install --upgrade "openmed[hf,service]"
uvicorn openmed.service.app:app --host 0.0.0.0 --port 8080

# or with Docker
docker build -t openmed:local .
docker run --rm -p 8080:8080 -e OPENMED_PROFILE=prod openmed:local

Example output:

INFO:     Uvicorn running on http://0.0.0.0:8080

curl -X POST http://127.0.0.1:8080/pii/extract \
  -H "Content-Type: application/json" \
  -d '{"text":"Paciente: Maria Garcia, DNI: 12345678Z","lang":"es"}'

Abbreviated example response:

{
  "text": "Paciente: Maria Garcia, DNI: 12345678Z",
  "entities": [
    {"text": "Maria Garcia", "label": "NAME", "confidence": 0.99, "start": 10, "end": 22},
    {"text": "12345678Z", "label": "ID", "confidence": 0.98, "start": 29, "end": 38}
  ],
  "model_name": "OpenMed/privacy-filter-multilingual"
}

Model lifecycle and service controls: free memory on demand with GET /models/loaded, POST /models/unload, and a keep_alive idle window; v1.8 also includes API-key/JWT auth, no-PHI request logging, tracing, gRPC, async jobs, webhooks, warm pools, dynamic batching, request coalescing, rate and concurrency limits, /livez, /readyz, and opt-in metrics:

OPENMED_SERVICE_KEEP_ALIVE=10m uvicorn openmed.service.app:app --host 0.0.0.0 --port 8080
curl -X POST http://127.0.0.1:8080/models/unload -H "Content-Type: application/json" -d '{"all":true}'

Example response:

{
  "unloaded": true,
  "released": {"models": 1, "tokenizers": 1, "pipelines": 1},
  "active_models": {}
}

See the full REST service guide.

Documentation

Full guides at openmed.life/docs.

AI agents can load the curated llms.txt index or the inlined llms-full.txt feed. Both are regenerated from the current documentation during every strict MkDocs build.


Getting Started	Analyze Text	Model Registry
FAQ	Anonymization	Batch Processing
Configuration Profiles	REST Service	MLX Backend
Transformers.js Export	FHIR Interop	HL7 v2 De-identification
OpenMed 1.9.1 Release Notes	OpenMed 1.9.0 Release Notes	Examples
Release Streams	Generative Model Policy	Contributing
Security Policy	Compliance Posture	Detector Plugin SDK
v1 to v2 Migration	MCP Client Connections	African Developer Onboarding

Meet the mascot

OpenMed's guardian is a fluffy Persian cat styled as a tiny Avicenna (Ibn Sina): the great Persian physician whose Canon of Medicine was the world's standard medical text for some 600 years. He keeps watch over the open book of medical knowledge, in a palette built around Persian turquoise (fīrūza): a local-first guardian for your most private data.

Contributing

Contributions welcome: bug reports, feature requests, and PRs alike. Please read the Contributing guide and our Code of Conduct first.

Open an issue
Contributing guide · Code of Conduct · Security policy
Translations welcome: help complete the other-language READMEs linked in the switcher at the top.

Security

Found a vulnerability? OpenMed redacts PHI, so a redaction bypass or PHI/PII leak is a security issue: please report it privately, never as a public issue. See SECURITY.md for the responsible-disclosure policy and the private reporting form. Never include real patient data in a report.

Credits

OpenMed builds on excellent open-source work: particular thanks to OpenAI (the Privacy Filter architecture), NVIDIA (the Nemotron PII dataset), Hugging Face (transformers, Transformers.js & the model ecosystem), Apple (MLX), and the Faker maintainers.

License

Released under the Apache-2.0 License. Third-party asset notices are recorded in NOTICE.

Citation

@misc{panahi2025openmedneropensourcedomainadapted,
      title={OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers for Biomedical NER Across 12 Public Datasets},
      author={Maziyar Panahi},
      year={2025},
      eprint={2508.01630},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.01630},
}

Expected result: BibTeX-compatible citation metadata for referencing OpenMed in papers, posters, and derived documentation.

Star History

If OpenMed is useful to you, a star helps others discover it.

Built by the OpenMed team

Website · Docs · X / Twitter · LinkedIn

Name		Name	Last commit message	Last commit date
Latest commit History 2,800 Commits
.claude-plugin		.claude-plugin
.claude		.claude
.github		.github
android		android
clients		clients
deploy		deploy
docs		docs
eval/redteam/corpus		eval/redteam/corpus
examples		examples
fixtures/parity		fixtures/parity
gates		gates
js		js
openmed		openmed
scripts		scripts
skills		skills
swift		swift
tests		tests
web/runtime		web/runtime
.dockerignore		.dockerignore
.editorconfig		.editorconfig
.gitignore		.gitignore
.gitleaks.toml		.gitleaks.toml
.pre-commit-config.yaml		.pre-commit-config.yaml
.secrets.baseline		.secrets.baseline
.swift-format		.swift-format
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
LICENSE		LICENSE
Makefile		Makefile
NOTICE		NOTICE
Package.resolved		Package.resolved
Package.swift		Package.swift
README.ar.md		README.ar.md
README.de.md		README.de.md
README.es.md		README.es.md
README.fa.md		README.fa.md
README.fr.md		README.fr.md
README.hi.md		README.hi.md
README.it.md		README.it.md
README.ja.md		README.ja.md
README.md		README.md
README.nl.md		README.nl.md
README.pt.md		README.pt.md
README.sw.md		README.sw.md
README.te.md		README.te.md
README.tr.md		README.tr.md
README.zh-CN.md		README.zh-CN.md
SECURITY.md		SECURITY.md
docker-compose.yml		docker-compose.yml
install-skills.sh		install-skills.sh
jitpack.yml		jitpack.yml
mkdocs.yml		mkdocs.yml
models.jsonl		models.jsonl
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Your Data. Your Model. Your Hardware.

See it in action

30-second example

Why OpenMed?

On-device on Apple: Swift, MLX & iOS

On-device on Android — Kotlin & ONNX Runtime Mobile

The same ONNX model on Python CPU

The same ONNX model in the browser

How it works

Quick start

Models

Privacy: PII detection & de-identification

Multilingual PII (26 supported languages)

REST API

Documentation

Meet the mascot

Contributing

Security

Credits

License

Citation

Star History

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Your Data. Your Model. Your Hardware.

See it in action

30-second example

Why OpenMed?

On-device on Apple: Swift, MLX & iOS

On-device on Android — Kotlin & ONNX Runtime Mobile

The same ONNX model on Python CPU

The same ONNX model in the browser

How it works

Quick start

Models

Privacy: PII detection & de-identification

Multilingual PII (26 supported languages)

REST API

Documentation

Meet the mascot

Contributing

Security

Credits

License

Citation

Star History

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages