Hulu-Med: A Transparent Generalist Model towards Holistic Medical Vision-Language Understanding

🔥 News

• [2026-04-01] 🚀 We released 🤗 Hulu-Med-Flash-Preview-27B. This version features shorter CoT reasoning for faster throughput and achieves an impressive 64.5 on HealthBench (43.7 on Hard set)! ⚡
• [2026-2-24] 🏆 We have released our 🤗 Hulu-Med-235A22 and 🤗 Hulu-Med-30BA3. Please have a try.
• [2025-11-27] ⚡ Hulu-Med is now compatible with the latest vLLM, offering faster inference and tensor parallel support! Thank you all for your patience and feedback 💪 see here for installation
• [2025-11-18] 🎊 We released Hulu-Med-4B, a lightweight model with strong multimodal and text reasoning abilities that surpasses MedGemma-4B and Lingshu-7B!
• [2025-11-01] 📊 Releasing our new evaluation code, MedUniEval! Built on MedEvalKit, MedUniEval is designed for the comprehensive evaluation of medical visual-language models across various modalities—including text, 2D, 3D, and video. More benchmarks are coming soon. Some processed evaluation data are available here.
• [2025-10-16] 🚀 Demo Is Live! We've just deployed a demo and we'd love for you to try it! Your insights and feedback are crucial for helping us improve the model in the next version.
• [2025-10-15] 🎉 Hulu-Med now supports Transformers integration! HuggingFace-compatible models released with simplified loading and inference. Integration with VLLM is ongoing. The HF models are now available in the main branch on Hugging Face.
• The model has been updated in the main branch of our Hugging Face repository. You can now load it directly using AutoModelForCausalLM.from_pretrained - the weights will be automatically downloaded. For users in regions with limited access, you can set the HF mirror environment variable to ensure reliable downloads:

export HF_ENDPOINT=https://hf-mirror.com

• [2025-10-08] Hulu-Med models and inference code released!

📖 Overview

Hulu-Med is a transparent medical vision-language model that unifies understanding across diverse modalities including medical text, 2D/3D images, and videos. Built with a focus on transparency and accessibility, Hulu-Med achieves state-of-the-art performance on 30 medical benchmarks while being trained entirely on public data.

Key Features

• 🌟 Holistic Multimodal Understanding: Seamlessly processes medical text, 2D images, 3D volumes, and surgical videos
• 🔓 Fully Transparent: Complete open-source pipeline including data curation, training code, and model weights
• 📊 State-of-the-Art Performance: Outperforms leading open-source models and competes with proprietary systems
• ⚡ Efficient Training: Only 4,000-40,000 GPU hours required for 7B-32B variants
• 🗂️ Comprehensive Coverage: Trained on 16.7M samples spanning 12 anatomical systems and 14 imaging modalities
• 🤗 Transformers Native: Now with native HuggingFace Transformers support for easier integration

Comprehensive Data Coverage

Our training corpus encompasses:

• 12 Major Anatomical Systems: Multi-System, Skin/Integumentary, Respiratory, Cellular/Tissue Level, Digestive, Nervous, Cardiovascular, Musculoskeletal, Reproductive, Urinary, Whole Body, Endocrine, Immune/Lymphatic, and Hematologic systems
• 14 Medical Imaging Modalities: CT, MRI, X-Ray, Ultrasound, PET, OCT, Endoscopy, Microscopy, Histopathology, Fundus, Dermoscopy, Angiography, Digital Photograph, and Medical Chart
• Diverse Downstream Tasks: Medical Dialogue, Anomaly Detection, Prognosis Prediction, Treatment Planning, Surgical Skill Assessment, Education, Medical Report Generation, Surgical Phase Recognition, Medical Computation, and more

🚀 Model Zoo

We provide three model variants with different parameter scales:

Note: HuggingFace-compatible versions (Hulu-Med-HF) are also available for easier integration with the Transformers library.

🏆 Performance Highlights

Medical Multimodal Benchmarks

Performance comparison on medical multimodal benchmarks (For the 'Medical VLM < 10B' subgroup, bold indicates the best method):

Medical Text Benchmarks

Performance comparison on medical text benchmarks (bold indicates the best method in each subgroup):

🛠️ Installation

# Clone the repository
git clone https://github.com/ZJUI-AI4H/Hulu-Med.git
cd Hulu-Med

# Create conda environment
conda create -n hulumed python=3.10
conda activate hulumed

# PyTorch and torchvision for CUDA 11.8
pip install torch==2.4.0 torchvision==0.19.0 --extra-index-url https://download.pytorch.org/whl/cu118

# Flash-attn pinned to a compatible version
pip install flash-attn==2.7.3 --no-build-isolation --upgrade

# Transformers and accelerate
pip install transformers==4.51.2 accelerate==1.7.0

# Video processing dependencies
pip install decord ffmpeg-python imageio opencv-python

# For 3D medical image processing (NIfTI files)
pip install nibabel

# Install other dependencies
pip install -r requirements.txt

🧩 vLLM Installation

VLLM_USE_PRECOMPILED=1 pip install git+https://github.com/jiangsongtao/vllm.git
pip install decord ffmpeg imageio
# No need to install requirements.txt

# or try this way
git clone https://github.com/jiangsongtao/vllm.git
cd vllm-main
export VLLM_USE_PRECOMPILED=1
rm -rf build/ .deps/
pip install -e .
pip uninstall flash-attn -y
pip install flash-attn --no-build-isolation

💻 Quick Start

We provide two ways to use Hulu-Med:

Option 1: Using HuggingFace Transformers (Recommended for Hulu-Med-HF models)

For easier integration, use the HuggingFace-compatible models with native Transformers support:

from transformers import AutoModelForCausalLM, AutoProcessor
import torch

model_path = "ZJU-AI4H/Hulu-Med-7B"

# Load model and processor
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
    attn_implementation="flash_attention_2",
)

processor = AutoProcessor.from_pretrained(
    model_path,
    trust_remote_code=True
)

tokenizer = processor.tokenizer

Text-Only Example

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Hello, I have a headache, what should I eat?"},
        ]
    }
]

modal = 'text'
inputs = processor(
    conversation=conversation,
    return_tensors="pt",
    add_generation_prompt=True
)

inputs = {k: v.to(model.device) if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        modals=[modal],
        temperature=0.6,
        max_new_tokens=4096,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

# Decode output
# Enable thinking mode by adding: "Please reason step by step, and put your final answer within \boxed{}."
# use_think=False: Only return the final answer without thinking process
# use_think=True: Include the model's reasoning/thinking process in the output
outputs = processor.batch_decode(
    output_ids,
    skip_special_tokens=True,
    use_think=False  # Set to True to see the thinking process
)[0].strip()
print(outputs)

2D Image Example

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": {
                    "image_path": "./demo/demo.jpg",
                }
            },
            {
                "type": "text",
                "text": "Generate a medical report for this image."
            },
        ]
    }
]

inputs = processor(
    conversation=conversation,
    add_system_prompt=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}

if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

output_ids = model.generate(**inputs, max_new_tokens=1024)
outputs = processor.batch_decode(
    output_ids,
    skip_special_tokens=True,
    use_think=False
)[0].strip()
print(outputs)

Multi Images Example

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "image", 
                "image": {
                    "image_path": "./demo/demo1.jpg",
                }
            },
            {
                "type": "image", 
                "image": {
                    "image_path": "./demo/demo2.jpg",
                }
            },
            {
                "type": "text", 
                "text": "Are these two images the same?"
            },
        ]
    }
]

inputs = processor(
    conversation=conversation,
    add_system_prompt=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

output_ids = model.generate(**inputs, max_new_tokens=1024)
outputs_no_think = processor.batch_decode(
    output_ids, 
    skip_special_tokens=True,
    use_think=False  
)[0].strip()
print(outputs_no_think)

Interleaved Example

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "text", 
                "text": "Image A:"
            },
            {
                "type": "image", 
                "image": {
                    "image_path": "./demo/XRay.jpg",
                }
            },
             {
                "type": "text", 
                "text": "Image B:"
            },
            {
                "type": "image", 
                "image": {
                    "image_path": "./demo/pathology.png",
                }
            },
            {
                "type": "text", 
                "text": "Which image is the pathology slide?"
            },
        ]
    }
]

inputs = processor(
    conversation=conversation,
    add_system_prompt=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

output_ids = model.generate(**inputs, max_new_tokens=1024)
outputs_no_think = processor.batch_decode(
    output_ids, 
    skip_special_tokens=True,
    use_think=False  
)[0].strip()
print(outputs_no_think)
#The pathology slide is Image B. It shows a microscopic view of tissue with various cellular structures and components, such as cells in different stages of maturation and areas of fibrous tissue. This type of image is typically used to examine the cellular architecture and identify any pathological changes within the tissue.

3D Medical Image Example

# Requires: pip install nibabel

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "3d",
                "3d": {
                    "image_path": "./demo/amos_0013.nii",
                    "nii_num_slices": 180,
                    "nii_axis": 2,  # 0=sagittal, 1=coronal, 2=axial
                }
            },
            {
                "type": "text",
                "text": "Generate a medical report for this 3D CT scan."
            },
        ]
    }
]

inputs = processor(
    conversation=conversation,
    add_system_prompt=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}

if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

output_ids = model.generate(**inputs, max_new_tokens=1024)
outputs = processor.batch_decode(
    output_ids,
    skip_special_tokens=True,
    use_think=False
)[0].strip()
print(outputs)

Video Example

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": {
                    "video_path": "./demo/1min_demo.mp4",
                    "fps": 1,
                    "max_frames": 1800
                }
            },
            {
                "type": "text",
                "text": "Describe this video in detail."
            },
        ]
    }
]

inputs = processor(
    conversation=conversation,
    add_system_prompt=True,
    add_generation_prompt=True,
    return_tensors="pt"
)

inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v 
          for k, v in inputs.items()}

if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

output_ids = model.generate(**inputs, max_new_tokens=1024)
outputs = processor.batch_decode(
    output_ids,
    skip_special_tokens=True,
    use_think=False
)[0].strip()
print(outputs)

Understanding the use_think parameter:

• use_think=False: Returns only the final answer (default for most use cases)
• use_think=True: Includes the model's internal reasoning/thinking process before the final answer

Option 2: Using Custom Loading (Original Method)

For the original Hulu-Med models (non-HF versions):

import torch
from hulumed import disable_torch_init, model_init, mm_infer
from hulumed.model import load_pretrained_model
from hulumed.mm_utils import load_images, process_images, load_video, process_video, tokenizer_multimodal_token, get_model_name_from_path, KeywordsStoppingCriteria
from hulumed.model.processor import HulumedProcessor
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
model_path = "path/to/your/model"
model_name = get_model_name_from_path(model_path)
tokenizer, model, image_processor, context_len = load_pretrained_model(
    model_path, None, model_name, device_map='cuda:0'
)
processor = HulumedProcessor(image_processor, tokenizer)

2D Example (Original Method)

slices = load_images("./demo/demo.jpg")
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "Describe this image in detail."},
        ]
    }
]
modal = 'image'
model = model.to("cuda:0")
inputs = processor(
    images=[slices] if modal != "text" else None,
    text=conversation,
    merge_size=2 if modal == "video" else 1,
    return_tensors="pt"
)
inputs = {k: v.cuda().to('cuda:0') if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        modals=[modal],
        temperature=0.6,
        max_new_tokens=8192,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)

3D Example (Original Method)

# We unify the modeling of video and 3D inputs as extensions along the temporal or spatial dimension
slices = load_images(
    "./demo/amos_0013.nii",  # Support NIfTI 3D input
    nii_num_slices=160
)
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "video", "num_frames": len(slices)},
            {"type": "text", "text": "This is a medical 3D scenario. Please generate a medical report for the given 3D medical images, including both findings and impressions."},
        ]
    }
]
modal = 'video'
model = model.to("cuda:0")
inputs = processor(
    images=[slices] if modal != "text" else None,
    text=conversation,
    merge_size=2 if modal == "video" else 1,
    return_tensors="pt"
)
inputs = {k: v.cuda().to('cuda:0') if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        modals=[modal],
        temperature=0.6,
        max_new_tokens=8192,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)

Video Example (Original Method)

frames, timestamps = load_video("./demo/1min_demo.mp4", fps=1, max_frames=3000)
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "video", "num_frames": len(frames)},
            {"type": "text", "text": "Please describe this video in detail."},
        ]
    }
]
modal = 'video'
model = model.to("cuda:0")
inputs = processor(
    images=[frames] if modal != "text" else None,
    text=conversation,
    merge_size=2 if modal == "video" else 1,
    return_tensors="pt"
)
inputs = {k: v.cuda().to('cuda:0') if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        modals=[modal],
        temperature=0.6,
        max_new_tokens=8192,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)

Text Example (Original Method)

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Hello, I have a headache, what should I do?"},
        ]
    }
]
modal = 'text'
model = model.to("cuda:0")
inputs = processor(
    text=conversation,
    merge_size=2 if modal == "video" else 1,
    return_tensors="pt"
)
inputs = {k: v.cuda().to('cuda:0') if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=True,
        modals=[modal],
        temperature=0.6,
        max_new_tokens=8192,
        use_cache=True,
        pad_token_id=tokenizer.eos_token_id,
    )

outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)

📊 Training

Data Preparation

Our training data consists of 16.7M samples across four categories:

• Medical Multimodal Data (9M samples): Covering 14 imaging modalities
• Medical Text Data (4.9M samples): Clinical notes, literature, QA pairs
• General Multimodal Data (1.3M samples): Enhancing generalization
• General Text Data (1.5M samples): Improving reasoning capabilities

Download and prepare the data: Coming soon

🏗️ Model Architecture

Hulu-Med consists of four core components:

1. Vision Encoder: SigLIP-based encoder with 2D RoPE for unified 2D/3D/video processing
2. Multimodal Projector: Projects visual tokens into language model space
3. LLM Decoder: Qwen-based decoder for generating responses
4. Medical-Aware Token Reduction: Efficient processing with ~55% token reduction

📋 Supported Tasks

• ✅ Visual Question Answering (2D/3D/Video)
• ✅ Medical Report Generation
• ✅ Disease Diagnosis
• ✅ Anatomical Understanding
• ✅ Surgical Phase Recognition
• ✅ Clinical Dialogue
• ✅ Medical Text Reasoning
• ✅ Multilingual Medical QA
• ✅ Rare Disease Diagnosis
• ✅ And more

📄 Citation

If you find Hulu-Med useful in your research, please cite:

@misc{jiang2025hulumedtransparentgeneralistmodel,
      title={Hulu-Med: A Transparent Generalist Model towards Holistic Medical Vision-Language Understanding}, 
      author={Songtao Jiang and Yuan Wang and Sibo Song and Tianxiang Hu and Chenyi Zhou and Bin Pu and Yan Zhang and Zhibo Yang and Yang Feng and Joey Tianyi Zhou and Jin Hao and Zijian Chen and Ruijia Wu and Tao Tang and Junhui Lv and Hongxia Xu and Hongwei Wang and Jun Xiao and Bin Feng and Fudong Zhu and Kenli Li and Weidi Xie and Jimeng Sun and Jian Wu and Zuozhu Liu},
      year={2025},
      eprint={2510.08668},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.08668}, 
}

@misc{jiang2025omnivmedscalingmedicalvisionlanguage,
      title={OmniV-Med: Scaling Medical Vision-Language Model for Universal Visual Understanding}, 
      author={Songtao Jiang and Yuan Wang and Sibo Song and Yan Zhang and Zijie Meng and Bohan Lei and Jian Wu and Jimeng Sun and Zuozhu Liu},
      year={2025},
      eprint={2504.14692},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.14692}, 
}

📜 License

This project is released under the Apache 2.0 License.

---