OpenMed-ZeroShot-NER-Protein-Multi-209M for OpenMed MLX

This repository contains an OpenMed MLX conversion of OpenMed/OpenMed-ZeroShot-NER-Protein-Multi-209M for Apple Silicon inference with OpenMed.

Artifact metadata:

• OpenMed MLX task: zero-shot-ner
• OpenMed MLX family: gliner-uni-encoder-span
• Weight format: safetensors
• Runtime API: GLiNERMLXPipeline

OpenMed MLX Status

• MLX rollout: refreshed for public access on 2026-06-23
• Hub artifact: OpenMed MLX repository
• Source checkpoint: OpenMed/OpenMed-ZeroShot-NER-Protein-Multi-209M
• Collection: OpenMed Medical MLX Models
• Runtime: OpenMed Python MLX backend on Apple Silicon
• Artifact layout: config.json, id2label.json, openmed-mlx.json, MLX weights, and tokenizer assets

Use This MLX Snapshot

Download this OpenMed MLX artifact directly from the Hub:

hf download OpenMed/OpenMed-ZeroShot-NER-Protein-Multi-209M-mlx --local-dir ./OpenMed-ZeroShot-NER-Protein-Multi-209M-mlx

Use the downloaded directory when you want to pin this exact MLX artifact in an offline or local Apple Silicon workflow.

Quick Start

pip install openmed
pip install "openmed[mlx]"

from huggingface_hub import snapshot_download
from openmed.mlx.inference import GLiNERMLXPipeline

model_path = snapshot_download("OpenMed/OpenMed-ZeroShot-NER-Protein-Multi-209M-mlx")
pipe = GLiNERMLXPipeline(model_path)

entities = pipe.predict_entities(
    "Patient John Doe was seen at Stanford Hospital.",
    labels=["person", "organization", "location"],
    threshold=0.5,
)

for entity in entities:
    print(entity)

Prompt packing metadata included with the model:

{
  "kind": "gliner-words",
  "entity_token": "<<ENT>>",
  "separator_token": "<<SEP>>",
  "class_token_index": 250103,
  "embed_marker_token": true,
  "split_mode": "words"
}

Swift and Apple Apps

Use Swift with OpenMedKit, not with MLX weight files directly.

1. Open Xcode and go to File > Add Package Dependencies.
2. Paste the OpenMed repository URL: https://github.com/maziyarpanahi/openmed
3. Choose the package product OpenMedKit from the repository.
4. Add a compatible CoreML model bundle plus id2label.json to your app target.

This MLX model is for Python services on Apple Silicon, local MLX inference on macOS, and Hub-hosted model distribution. If a given environment cannot write weights.safetensors, OpenMed falls back to weights.npz so the model remains usable.

Credits

• Base checkpoint: OpenMed/OpenMed-ZeroShot-NER-Protein-Multi-209M
• OpenMed GitHub: https://github.com/maziyarpanahi/openmed
• OpenMed website: https://openmed.life
• MLX conversion and runtime support: OpenMed
• Swift runtime for Apple apps: OpenMedKit from the OpenMed repository