Stripping identifiers from a clinical record directly on an Android phone, in airplane mode, without a single byte leaving the device — that's what OpenMed 1.8 does, released this week under Apache 2.0. The clinical NLP toolkit follows a hard rule: patient data never leaves the hardware it's processed on. No cloud calls, no API keys. That architectural stance becomes concrete on three new fronts: OpenMedKit for Android, iOS/Swift and React Native bridges, and a browser runtime accelerated by WebGPU.

OpenMedKit leverages ONNX Runtime Mobile and ML Kit OCR to scan a document and remove names, medical record numbers, and dates entirely on the phone. On the browser side, Transformers.js and ONNX Runtime Web (wasm + WebGPU backends) perform the same work fully client-side, eliminating the server's attack surface. For a radiology department dealing with DICOM files, de-identification even reaches burned-in pixel text via OCR — a detail that defeats many superficial “redactions.”

The PDF redaction scam

One of the release's most instructive tools is verify-pdf. Many “redacted” PDFs merely draw a black box over the text: the underlying text layer survives intact, and a simple copy-paste pulls the name right out. OpenMed deliberately fails validation unless the text is truly gone, forcing developers to confront the brittleness of current practices. It's a technical reminder: GDPR compliance isn't satisfied with visual tricks but requires the certain destruction of identifying information.

The global lab of version 1.9

The next iteration is already being built in the open, with over 400 issues. The maintainer invites the community to contribute language packs with national-ID validators: from the Estonian isikukood to the Serbian JMBG, the Croatian OIB, and the Bulgarian EGN. New clinical NER domains — pediatric growth parameters, spirometry, immunizations — are also on the table, along with RTF and ODT text extraction with character-offset maps. Eleven outside contributors shipped code in 1.8, many having started with exactly these sorts of issues, signaling that the project is growing a distributed, self-sustaining developer base.

Looking at the models, the OpenMed ecosystem counts over 1,500 on Hugging Face, all Apache 2.0. The PII models run on MLX (Apple Silicon), GGUF/llama.cpp, ONNX, or plain transformers. Two of them hold the first and second spots on the independent PII Masking Benchmark English leaderboard, and the 44-million-parameter model — ranked fourth — is small enough for a phone. The message is clear: local processing is no compromise. Quality is there, coupled with zero external data traffic.

For those managing healthcare infrastructure, this paradigm shift has deep economic and operational implications. A de-identification pipeline that runs entirely on enterprise hardware or edge devices moves spending from an operational model (per-API-call cost) to a capital investment, reducing lock-in risk and simplifying compliance reporting. In air-gapped settings — think genetic research labs or clinical trials — the “local by design” approach becomes the only viable path. OpenMed’s commitment to remaining fully self-hosted and cloud-dependency-free aligns with the broader industry trend toward data sovereignty by default, accelerated by stricter regulations and the awareness that health data is too valuable a target to entrust to third parties.