A 66M DistilBERT model for extracting anatomy mentions in English clinical and biomedical text.
Runs locally after download on Python CPU, in the browser, and on Android.

OpenMed | Documentation | Model collection | Paper

Model

Field	Value
Task	Anatomy named-entity recognition
Language	English
Architecture	DistilBERT
Parameters	66M (66,000,000)
Maximum sequence length	512 tokens
Entity labels	`Anatomy`
Source model	`OpenMed/OpenMed-NER-AnatomyDetect-TinyMed-66M`
License	apache-2.0

OpenMed in Python on CPU

pip install --upgrade "openmed[onnx-runtime]"

python

from openmed import OnnxModel

model = OnnxModel.from_pretrained("OpenMed/OpenMed-NER-AnatomyDetect-TinyMed-66M-v1-onnx-android")
entities = model("The biopsy was taken from the left lung.")

for entity in entities:
    print(entity.to_dict())

OpenMed selects the CPU-oriented INT8 graph by default and returns labels, confidence scores, exact character offsets, and source text.

OpenMed in Web

npm install openmed @huggingface/transformers onnxruntime-web

typescript

import { loadOnnxModel } from "openmed";

const repo = "OpenMed/OpenMed-NER-AnatomyDetect-TinyMed-66M-v1-onnx-android";
const model = await loadOnnxModel(repo);
const entities = await model("The biopsy was taken from the left lung.");

The default INT8 path runs with WebAssembly. For WebGPU, select the FP16 graph:

typescript

const model = await loadOnnxModel(repo, {
  variant: "fp16",
  device: "webgpu",
});

OpenMedKit for Android

Add JitPack to the consumer application's settings.gradle.kts:

kotlin

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

Use the latest OpenMed build from the master branch:

kotlin

dependencies {
    implementation("com.github.maziyarpanahi:openmed:master-SNAPSHOT")
}

After downloading this model repository into an app-controlled directory:

kotlin

import com.openmed.openmedkit.OpenMedKit

OpenMedKit.fromDirectory(modelDirectory).use { model ->
    val entities = model.analyzeText("The biopsy was taken from the left lung.")
}

Inference and tokenization remain on-device.

Included Artifacts

Artifact	Recommended use
`model_int8.onnx`	CPU, WebAssembly, and Android default
`model_fp16.onnx`	WebGPU and compatible accelerated runtimes
`model.onnx`	Full-precision reference
`model.ort`	Custom ONNX Runtime Mobile integration
`tokenizer.json`	Cross-platform tokenizer
`openmed-onnx.json`	Runtime contract and operator metadata

All graphs use opset 18, dynamic batch and sequence axes, stable tensor names, and source-text offset metadata.

The OpenMed Ecosystem

This model is part of OpenMed, an Apache-2.0, local-first clinical AI stack:

2,000+ medical models for clinical NER, biomedical extraction, and privacy.
PII detection and de-identification across 55+ identifier types and 17 languages.
Python, MLX, Swift, Android, React Native, Web, REST, and gRPC runtimes.
Structured and multimodal intake for OCR, documents, DICOM, FHIR, and HL7.
Offline and air-gapped deployment with no telemetry by default.

Intended Use and Limitations

This model is intended for extracting anatomy mentions. It does not diagnose conditions or make clinical decisions. Evaluate recall, thresholds, and span behavior on appropriately governed data before deployment. Local execution supports privacy-preserving workflows but does not by itself guarantee regulatory compliance.

All examples in this card are synthetic.

Citation

bibtex

@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}
}