MedMO-8B-Next is the latest and most powerful iteration of the MedMO family — an open-source multimodal foundation model purpose-built for comprehensive medical image understanding and grounding. Trained on 26M+ diverse medical samples across 45 datasets, MedMO-8B-Next achieves state-of-the-art performance across all major medical imaging benchmarks, outperforming both open-source and closed-source competitors on VQA, Text QA, grounding, and report generation tasks.
MedMO excels at a comprehensive range of medical imaging tasks:
- Visual Question Answering (VQA): Answer complex questions about medical images across radiology, pathology, ophthalmology, and dermatology
- Text-Based Medical QA: Clinical reasoning and medical knowledge question answering
- Radiology Report Generation: Generate detailed, clinically accurate radiology reports from medical images
- Disease Localization with Bounding Boxes: Precise spatial detection and localization of pathological findings
- Anatomical Grounding: Spatial localization and grounding of anatomical structures
- Clinical Reasoning: Step-by-step diagnostic reasoning and clinical decision support
- Diagnostic Classification: Multi-class disease classification across diverse imaging modalities
- Spatial Object Detection: Fine-grained detection in microscopy, pathology slides, and cellular imaging
- Medical Report Summarization: Extract and summarize key clinical findings from complex medical reports
- Radiology (X-ray, CT, MRI, Ultrasound)
- Pathology & Microscopy
- Ophthalmology (Fundus, OCT)
- Dermatology
- Nuclear Medicine (PET, SPECT)
MedMO-8B-Next sets a new state-of-the-art across the board, achieving the highest average scores on both medical VQA and Text QA benchmarks — surpassing strong baselines including Lingshu-7B and Fleming-VL-8B.
OMIVQA = OmniMedVQA · MedXQA = MedXpertQA · Medbullets reported as op4/op5
| Model | MMMU-Med | VQA-RAD (closed/all) | SLAKE (closed/all) | PathVQA | PMC-VQA | OmniMedVQA | MedXpertQA | Avg. |
|---|---|---|---|---|---|---|---|---|
| Lingshu-7B | 54.0 | 77.2 / 43.0 | 82.4 / 33.2 | 41.9 | 54.2 | 82.9 | 26.9 | 55.1 |
| Fleming-VL-8B | 63.3 | 78.4 / 56.4 | 86.9 / 80.0 | 56.5 | 64.3 | 88.2 | 21.6 | 66.1 |
| MediX-R1-8B | 63.3 | 75.2/51.6 | 70.3/54.4 | 41.0 | 55.3 | 73.8 | 24.9 | 57.1 |
| MedMO-4B | 54.6 | 50.9 / 35.0 | 41.0 / 30.0 | 42.4 | 50.6 | 79.7 | 24.8 | 45.4 |
| MedMO-8B | 64.6 | 72.3 / 64.7 | 70.6 / 70.0 | 56.3 | 59.4 | 84.8 | 26.2 | 63.2 |
| MedMO-4B-Next | 58.7 | 79.7 / 59.6 | 78.0 / 74.0 | 73.3 | 75.7 | 90.6 | 27.0 | 68.5 |
| MedMO-8B-Next | 69.3 | 86.4 / 68.0 | 83.0 / 81.6 | 56.3 | 74.1 | 93.3 | 42.9 | 72.7 |
| Model | MMLU-Med | PubMedQA | MedMCQA | MedQA | Medbullets (op4/op5) | MedXpertQA | SGPQA | Avg. |
|---|---|---|---|---|---|---|---|---|
| Lingshu-7B | 69.6 | 75.8 | 56.3 | 63.5 | 62.0 / 53.8 | 16.4 | 27.5 | 53.1 |
| Fleming-VL-8B | 71.8 | 74.0 | 51.8 | 53.7 | 40.5 / 37.3 | 12.1 | 24.9 | 45.7 |
| MediX-R1-8B | 79.0 | 73.4 | 60.1 | 85.8 | 55.1/47.0 | 14.4 | 34.3 | 56.1 |
| MedMO-4B | 75.7 | 78.0 | 58.0 | 78.5 | 57.5 / 47.7 | 16.4 | 29.4 | 55.1 |
| MedMO-8B | 81.0 | 77.6 | 65.0 | 84.3 | 66.5 / 60.2 | 19.9 | 36.0 | 61.3 |
| MedMO-4B-Next | 74.8 | 78.2 | 58.1 | 78.3 | 57.4 / 47.6 | 16.5 | 29.5 | 55.0 |
| MedMO-8B-Next | 80.2 | 75.6 | 62.0 | 83.8 | 65.2 / 57.8 | 20.9 | 35.5 | 60.1 |
Bold = best result, underline = second-best result.
- Benchmarked on AMD MI210 GPU.
pip install transformers torch qwen-vl-utilsfrom transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
# Load model
model = Qwen3VLForConditionalGeneration.from_pretrained(
"MBZUAI/MedMO-8B",
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map="auto",
)
processor = AutoProcessor.from_pretrained("MBZUAI/MedMO-8B")
# Prepare your input
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": "path/to/medical/image.png",
},
{"type": "text", "text": "What abnormalities are present in this chest X-ray?"},
],
}
]
# Process and generate
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to(model.device)
# Generate output
generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text[0])messages = [
{
"role": "user",
"content": [
{"type": "image", "image": "chest_xray.png"},
{"type": "text", "text": "Detect and localize all abnormalities in this image."},
],
}
]
# Output: "Fractures <box>[[156, 516, 231, 607], [240, 529, 296, 581]]</box>"messages = [
{
"role": "user",
"content": [
{"type": "image", "image": "ct_scan.png"},
{"type": "text", "text": "Generate a detailed radiology report for this CT scan."},
],
}
]
# MedMO generates comprehensive clinical reports with findings and impressionsMedMO is built on Qwen3-VL-8B-Instruct and trained through a 4-stage progressive pipeline:
- Stage 1 - General Medical SFT: Large-scale training on 18.5M image-text pairs for foundational medical understanding
- Stage 2 - High-Resolution & Grounding: Training on 3M curated samples at 1280×1280 resolution for spatial localization
- Stage 3 - Instruction Tuning: Fine-tuning on 4.3M instruction-response pairs for task-specific alignment
- Stage 4 - Reinforcement Learning: GRPO training with verifiable rewards (label accuracy, bbox IoU) for enhanced grounding
Total Training Data: 26M+ samples from 45 medical datasets spanning diverse modalities and anatomical systems.
MedMO was comprehensively evaluated using multiple frameworks and metrics:
We used MedEvalKit for systematic evaluation across all QA taks:
For VQAs question answering and report generation tasks, we employed an LLM-as-a-Judge methodology:
- Judge Model: gpt-5-mini-2025-08-07
This approach provides robust, scalable assessment of the model's ability to generate detailed, medically accurate long-form responses.
| Model | Parameters | Best For |
|---|---|---|
| MedMO-8B-Next | 8B | SOTA highest accuracy, all tasks — recommended |
| MedMO-4B-Next | 4B | 2nd SOTA, high accuracy in resource-constrained environments |
| MedMO-8B | 8B | Previous generation |
| MedMO-4B | 4B | Resource-constrained environments |
For detailed benchmark results, please refer to our paper.
If you use MedMO in your research, please cite our paper:
@article{deria2026medmo,
title={MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images},
author={Deria, Ankan and Kumar, Komal and Dukre, Adinath Madhavrao and Segal, Eran and Khan, Salman and Razzak, Imran},
journal={arXiv preprint arXiv:2602.06965},
year={2026}
}We gratefully acknowledge the following:
- Base Architecture: Built on Qwen3-VL by Alibaba Cloud
- Evaluation Framework: MedEvalKit by Alibaba DAMO Academy
- Training Framework: TRL (Transformer Reinforcement Learning) by Hugging Face
- LLM-as-a-Judge: Evaluation powered by gpt-5-mini-2025-08-07 from OpenAI
- Compute Resources: Training conducted on 64× AMD Instinct MI210 GPUs
- Open-Source Datasets: We thank the medical imaging community for providing high-quality public datasets including MedTrinity, MIMIC-CXR, CheXpert, PathVQA, and many others that made this work possible
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
For questions, issues, or collaborations:
- Email: ankan.deria@mbzuai.ac.ae