OpenMed Hits 6 Million Downloads: One Developer's Open-Source Medical AI Stack

By Prahlad Menon Published 2026-06-10 2 min read

Six million downloads. Over a thousand open medical AI models. Twenty-plus curated datasets. Apache 2.0 licensed. Built by a single developer without external funding.

OpenMed just shipped v1.5.5, and the numbers tell an unusual story: someone built the largest open-source medical AI ecosystem while the rest of the industry was chasing cloud API margins.

What OpenMed Actually Is

OpenMed isn’t a chatbot. It’s specialized encoder transformers — BERT, ELECTRA, DeBERTa families — fine-tuned for biomedical named-entity recognition. Pull structured entities out of unstructured clinical text: diseases, drugs, genes, anatomy, oncology terms. The models stay small enough to run on a laptop or phone.

The signature capability: PII detection and de-identification across 12 languages. All 18 HIPAA Safe Harbor identifiers. Locale-aware validators for SSN, NIR, Steuer-ID, Codice Fiscale, DNI, BSN, CPF/CNPJ. Smart entity merging that keeps fragmented tokens like 123-45-6789 intact instead of split across subwords.

And critically: everything runs on-device. No cloud calls. No vendor lock-in. No patient data leaving your network.

from openmed import analyze_text

result = analyze_text(
    "Patient started on imatinib for chronic myeloid leukemia.",
    model_name="disease_detection_superclinical",
)
# DISEASE: chronic myeloid leukemia (0.98)
# DRUG: imatinib (0.95)

The 1.5.5 Release

The latest release adds:

Batch PII extraction and de-identification — up to 3.3× higher throughput on CPU, 2.2× on MLX vs. single-document processing
Model lifecycle controls for self-hosting: load, unload, keep-alive idle windows
Chunked PII processing for long clinical documents in Swift/OpenMedKit
Documentation in 13 languages

The batch processing is significant. Healthcare systems don’t process one clinical note at a time — they process thousands. The new BatchProcessor with operation="extract_pii" or "deidentify" lets you parallelize that workload properly.

The MLX Story

OpenMed runs natively on Apple Silicon via MLX — 24-33× faster than CPU PyTorch for the Privacy Filter models. Same model name works across platforms: on non-Apple hardware, MLX model names automatically fall back to the matching PyTorch checkpoint.

For iOS/macOS apps, OpenMedKit brings the same PII detection and clinical extraction into native Swift:

dependencies: [
    .package(url: "https://github.com/maziyarpanahi/openmed.git", from: "1.5.5"),
]

On-device PII redaction in a native app, offline, no PHI leaving the device. That’s the play.

The Model Catalog

The sheer breadth is impressive:

Model	Specialization	Size
`disease_detection_superclinical`	Diseases & conditions	434M
`pharma_detection_superclinical`	Drugs & medications	434M
`pii_superclinical_large`	PII & de-identification	434M
`anatomy_detection_electramed`	Anatomy & body parts	109M
`gene_detection_genecorpus`	Genes & proteins	109M

Plus 247 PII checkpoints across 12 languages. And the Privacy Filter family — three variants on the OpenAI Privacy Filter architecture (sparse-MoE transformer with local attention, sink tokens, RoPE+YaRN, tiktoken o200k_base):

Original OpenAI Privacy Filter
Nemotron-PII fine-tune (NVIDIA’s PII dataset)
OpenMed Multilingual

All available in PyTorch and MLX, including 8-bit quantized variants.

Why This Matters

The broader trend is worth watching: healthcare AI is moving toward open, local, and self-hosted infrastructure rather than black-box APIs.

HIPAA doesn’t care that your cloud vendor signed a BAA — data breaches happen, and the liability cascades. Air-gapped deployments, on-prem inference, edge devices — these aren’t edge cases in healthcare. They’re the compliance-friendly path.

OpenMed represents what happens when someone actually builds for that reality instead of treating it as an afterthought. Your choice of model. Your laptop. Your cloud. Your on-prem deployment. Zero vendor lock-in.

The fact that this entire stack — 1,000+ models, 20+ datasets, MLX acceleration, Swift native apps — was built by a single developer (Maziyar Panahi) without external funding makes it more impressive. And more sustainable, arguably. No VC timeline, no pivot pressure, no acqui-hire exit that orphans the project.

Getting Started

# Core + Hugging Face runtime
pip install "openmed[hf]"

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

# Apple Silicon acceleration
pip install "openmed[mlx]"

The documentation at openmed.life/docs is thorough. The PII Detection Complete Guide notebook walks through the full workflow.

For production REST deployments:

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

Air-gapped? Point model_id at a local directory:

from openmed import OpenMedConfig, analyze_text

result = analyze_text(
    "Patient presents with chronic myeloid leukemia.",
    model_id="./models/OpenMed-NER-DiseaseDetect-SuperClinical-434M",
    config=OpenMedConfig(device="cpu"),
)

Six million downloads in under a year. The open-source medical AI stack is real, it runs locally, and it’s Apache 2.0. That’s the story.

Links: