Medical NLP that runs in a Swift app, not a $50k contract

What it does OpenMed is a Python toolkit and Swift package for extracting structured medical entities from clinical text. It bundles a registry of 12+ specialized transformer models—disease detection, drug identification, anatomy, genes, and PII—behind a single analyze_text() call. A FastAPI service and batch processor wrap the same models for backend use, while OpenMedKit ports the PII and privacy-filter workloads to MLX on macOS and iOS.

The interesting bit The project treats “production-ready” as a feature, not a slogan. It ships configuration profiles (dev/prod/fast), request validation with unified error envelopes, and smart entity merging to fix tokenization fragmentation on dates and clinical IDs. The de-identification layer is unusually thorough: six masking strategies including Faker-backed locale-aware surrogates, cryptographic hashing, and deterministic replacement with per-document consistency.

Key highlights

• Runs fully offline with local model paths; explicitly blocks Hugging Face Hub calls when a local directory is supplied
• Apple Silicon acceleration via both Python MLX and Swift OpenMedKit; CoreML fallback for bundled models
• Privacy Filter family re-implements OpenAI’s sparse-MoE architecture (local attention, sink tokens, RoPE+YaRN) on three training sets: OpenAI’s own, NVIDIA Nemotron-PII, and a multilingual variant
• PII coverage claims all 18 HIPAA Safe Harbor identifiers with configurable confidence thresholds
• Apache 2.0 licensed; Docker image and Colab notebook provided

Caveats

• The broader OpenMed model-to-Swift conversion pipeline is still “active work,” so only PII, Privacy Filter, and experimental GLiNER tasks are currently portable to iOS
• README claims models “outperform proprietary solutions” but cites no benchmarks or comparison numbers
• Most models are in the 109M–434M range; the “12+” registry is real but the catalog table only shows five entries

Verdict Worth a look if you need clinical NER or de-identification without cloud lock-in, especially in Apple-ecosystem deployments. Skip if you require validated accuracy claims against commercial competitors or mature Android/Windows on-device support.