OpenMed 2.0.0 · on-device clinical AI
OpenMed reads clinical text and removes 55+ PHI types on the hardware you control, so patient data never leaves the device. 2,000+ Apache-2.0 models, 33 model-backed PII languages, state of the art on 10 of 12 biomedical NER benchmarks — and a new release most weeks.
Model downloads · all-time
340M
One person on lunch breaks in July 2025; the largest open medical-AI collection by July 2026. Founded by Maziyar Panahi. Join on GitHub
uv pip install "openmed[hf]"
# local extraction + de-identification — nothing leaves the process
from openmed import deidentify, extract_pii
note = "Patient Casey Example (MRN 00123) called from 555-0100."
result = extract_pii(note) # entities + offsets + confidence
print(deidentify(note))
# → "Patient [NAME] ([ID]) called from [PHONE]."No API key to provision, no procurement call, no data-processing agreement to negotiate. Install the package and the same call shape runs on a laptop CPU, Apple Silicon, or a GPU server.
analyze_text(...)
One-call inference with structured outputs
extract_pii / deidentify
Language-aware across 34 supported PII languages
BatchProcessor(...)
Up to 3.3× throughput on CPU, 2.2× on MLX
openmed.service.app
REST and gRPC, auth and no-PHI logging built in
Cloud medical APIs want your notes. Licensed clinical NLP wants a per-server subscription. Research toolkits stopped shipping. OpenMed is the only one of the four that runs on a phone, costs nothing, and shipped this week.
| Capability | OpenMed Apache-2.0 · on-device | Cloud APIs AWS Comprehend Medical · Azure Health · Google Healthcare NLP | Licensed clinical NLP Enterprise NLP suites, per-server subscription | Research toolkits medspaCy · MedCAT · cTAKES |
|---|---|---|---|---|
| Where inference runs The question every compliance review starts with | Your hardware — phone to GPU server | Vendor cloud region | Your servers, per-server licence | Your servers |
| Patient data leaves the network | Never | Sent to the vendor | No | No |
| What it costs at 10M notes | $0 — Apache-2.0, weights included | Per-character billing | Per-server subscription | $0 |
| Languages supported PII and clinical coverage, per stack | 33 model-backed · 34 supported codes | Mostly English endpoints | A few clinical locales, licensed | English only |
| Runs on iPhone and Android Where clinicians actually stand | OpenMedKit · ONNX Runtime Mobile | — | — | — |
| Apple Silicon acceleration | MLX — 24–33× faster than CPU | n/a | — | — |
| Runs in the browser | Transformers.js · WebGPU | — | — | — |
| CPU-optimised ONNX builds Most hospital hardware has no GPU | fp32, fp16 and INT8 exports, ORT Mobile ready | n/a — cloud only | Heavyweight JVM runtime | CPU, unoptimised |
| Release cadence How fast a gap gets closed | Weekly, in public | Vendor roadmap | Quarterly majors | Largely dormant |
| Benchmarks you can rerun | arXiv:2508.01630 · open weights + scripts | Vendor white papers | Vendor-run comparisons | Academic papers |
Capability and cadence rows describe publicly documented positions · July 2026 Sources: vendor docs · arXiv:2508.01630 · OpenMed release feed
Same model names, five runtimes. iPhone, iPad and Android through OpenMedKit, Apple Silicon through MLX, plain CPUs and NVIDIA GPUs through PyTorch, the browser through Transformers.js and WebGPU, and REST or gRPC services in your own VPC.
Apple Silicon · MLX
Median latency per inference step on the Privacy Filter
family. Install openmed[mlx] and a MacBook
becomes the inference tier.
OpenMedKit · Swift + Kotlin
Scan a note on an iPhone, redact it, extract clinical signals — offline, with ONNX Runtime Mobile on Android and Kotlin parity tests on span boundaries.
Explore OpenMedKit docsBrowser · WebGPU
Export ONNX token-classification artifacts into the
transformersjs/ layout and redact in the
tab, before submit.
All 18 HIPAA Safe Harbor identifiers inside a 55+ entity catalog: 33 model-backed PII languages, 34 supported codes, 600+ PII models, plus validator-backed ID-only locales. Policy-aware pipelines add calibrated thresholds, signed audit reports, re-identification risk scoring, and release gates that fail closed when the evidence is missing. 34 supported PII language codes: am, ar, as, bn, cs, da, de, el, en, es, fr, he, hi, id, it, ja, ko, mr, nl, no, or, pt, ro, ru, sv, sw, ta, te, th, tr, uk, xh, zh, and zu
Every run produces canonical span records and a signed report: what was found, what policy applied, what was released. Release gates fail closed.
audit-trailrisk-scoringPresidio-inspired scoring boosts confidence when keywords like SSN:, DOB: or NPI: appear near a detected entity — the framework's idea, trained on clinical language.
keyword-boosting100-char-windowBuilt-in validators cut false positives: French NIR, German Steuer-ID, Italian Codice Fiscale, Spanish DNI/NIE, Dutch BSN, CPF/CNPJ, Luhn.
NIRCodice-FiscaleCPF/CNPJSubword tokenizers split 123-45-6789 into fragments; semantic patterns reassemble complete SSNs, phone numbers and multi-word entities before redaction.
BIO-tagsspan-parityMask with [NAME] labels, redact, hash, shift dates by a fixed offset, or replace with Faker-backed surrogates that keep local ID formats valid.
maskreplaceshift_datesNo API calls, no licence check-in, no telemetry. Point a model name at a local directory and OpenMed never contacts the Hub — air-gapped by configuration, not by promise.
air-gappedno-telemetrySpecialised models across 13 biomedical domains — chemicals, diseases, genes, proteins, species, anatomy, oncology — sized from 33M to 568M so the small ones fit on a phone.
Showing all six model examples.
ElectraMed · 33M
BioMed · 335M
PubMed · 109M
MultiMed · 568M
BigMed · 278M
SuperClinical · 184M
The library is free and stands alone. Welna and OpenMed Agent are separate products built on it — one for patients, one for clinical teams.
Reads Apple Health with your permission, redacts identifiers on device, and turns ninety days of signals into plain-language reads, plus the questions worth bringing to a clinician.
Runs the OpenMed PII model on device (~40 MB), redacting identifiers before anything leaves the phone.
Download on the App StoreOne inspectable agent for prior auth, appeals, coding, claims, care coordination and FHIR work, with visible plans and auditable artifacts before any action.
Built on OpenMed extraction, de-ID & terminology.
Open agent.openmed.lifePaper · arXiv:2508.01630
Domain-adaptive pre-training with parameter-efficient LoRA on 350k biomedical passages set state of the art on 10 of 12 public NER benchmarks. Weights, recipes and eval scripts are all published, and a full fine-tune finishes in under 12 hours on a single GPU.
If yours isn't here, open an issue — you'll get an answer from the people who train the models.
Cloud APIs bill per character and require sending clinical text to a vendor region — every new use case becomes a privacy review. OpenMed downloads once and runs in your process: laptop, VPC, or air-gapped box, with no telemetry and no licence check-in. It is also the only option here that runs inside an iPhone or a browser tab.
Those are mature platforms with strong published accuracy and support contracts — if you need a vendor SLA, that is a real answer. OpenMed competes on ownership: no per-server subscription, Apache-2.0 weights and recipes, mobile and browser runtimes, and a release most weeks instead of quarterly. Our benchmarks you can rerun; theirs you take on trust.
Yes. Every model ships with its full training recipe: domain-adaptive pre-training plus LoRA, updating under 1.5% of parameters. A full run finishes in under 12 hours on a single GPU (under 1.2 kg CO2e), and starter notebooks live in the openmed-starter repo.
Releases are versioned, signed and documented, with pinned model artifacts, calibrated thresholds and release gates that fail closed. Pin a version and upgrade on your schedule — the cadence just means a gap you report closes in days, not in the next procurement cycle.
Language packs, national-ID validators, translations and mobile runtimes have all come from people outside the original author. Issues and PRs are public, security disclosure is documented, and PHI-leak reports get a private channel. If your locale is missing, requesting it is a realistic path to it existing.
No. They are compact encoder transformers fine-tuned for token classification — they label spans of text rather than generating it, so there is nothing to hallucinate. Small enough to run on a phone; the Privacy Filter family adds the de-identification policy layer on top.
Open source, open invitation
Every model, validator and language pack came from someone who needed it and shipped it in public. Star the repo, request a language, file a redaction bypass, send a pull request — clinical AI gets better when more people can see how it works.