MedMLX/biomedparse-v2-mlx

BioMedParse v2

Segment 3D medical images with text prompts using Microsoft BioMedParse v2 and native MLX.

GitHub · Files · Docs · Paper

Input	Output
Medical image + text prompt	Prompted segmentation

Apple Silicon · Metal · Python 3.13

Converted from the Microsoft BioMedParse v2 checkpoint; no retraining was performed.

Quick start

uv venv --python 3.13
source .venv/bin/activate

uv pip install "biomedparse-mlx>=0.1.5"
uv tool install huggingface-hub

hf download MedMLX/biomedparse-v2-mlx --revision 1c251c6561aa143cd79ff0ccdfa53a1a6e60c7f0 --local-dir models

Replace the example input paths with your own files.

biomedparse-mlx segment \
  --image ct.nii.gz --modality ct --ct-site abdomen \
  --weights models/biomedparse-v2.safetensors \
  --tokenizer models/tokenizer.json \
  --prompt 'Visualization of the liver in abdominal CT imaging' \
  --output-dir outputs/liver --seed 0

Select the CT site for your anatomy. The example uses abdominal CT.

Use & license

For research and education. Not validated for clinical use.

Weights: CC-BY-NC-SA-4.0. Use is noncommercial and adaptations must retain the same license. The original Microsoft checkpoint is intended for research and development, not clinical use. MedMLX converted the checkpoint to native MLX/safetensors without retraining.

The included tokenizer is from OpenAI CLIP. See NOTICE and upstream notices for attribution and terms.

Paper references

• Original model: BiomedParse: a biomedical foundation model for image parsing.
• BiomedParse v2 architecture: Boltzmann Attention Sampling for Image Analysis with Small Objects.

Citation

If you use MedMLX, please cite the MedMLX paper:

@article{sandoval2026medmlx,
  title   = {MedMLX: On-Device Medical Imaging AI for Apple Silicon},
  author  = {Sandoval, A. and Sandoval, J. and Saleem, A.},
  journal = {medRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.28.26366762}
}