MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images

April 8, 2026 · View on GitHub

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.

🎯 Capabilities

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

Supported Modalities

Radiology (X-ray, CT, MRI, Ultrasound)
Pathology & Microscopy
Ophthalmology (Fundus, OCT)
Dermatology
Nuclear Medicine (PET, SPECT)

🏆 Benchmark Performance

VQA & Text QA Results

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

Medical VQA Benchmarks

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

Medical Text QA Benchmarks

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.

🚀 Quick Start

Installation

pip install transformers torch qwen-vl-utils

Basic Usage

from 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])

Example: Disease Localization with Bounding Boxes

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>"

Example: Report Generation

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 impressions

🏗️ Model Architecture

MedMO 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.

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 Family

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.

📄 Citation

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}
}

🙏 Acknowledgments

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

📜 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

📧 Contact

For questions, issues, or collaborations:

Email: ankan.deria@mbzuai.ac.ae