I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.

In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp

While this does increase model file size, it significantly improves precision for a given quantization level.

I'd love your feedback—have you tried this? How does it perform for you?

Click here to get info on choosing the right GGUF model format

Website 🤖 7B Model 🤖 32B Model MedEvalKit Technical Report

Lingshu - SOTA Multimodal Large Language Models for Medical Domain

BIG NEWS: Lingshu is released with state-of-the-art performance on medical VQA tasks and report generation.

This repository contains the model of the paper Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning. We also release a comprehensive medical evaluation toolkit in MedEvalKit, which supports fast evaluation of major multimodal and textual medical tasks.

Highlights

Lingshu models achieve SOTA on most medical multimodal/textual QA and report generation tasks for 7B and 32 model sizes.
Lingshu-32B outperforms GPT-4.1 and Claude Sonnet 4 in most multimodal QA and report generation tasks.
Lingshu supports more than 12 medical imaging modalities, including X-Ray, CT Scan, MRI, Microscopy, Ultrasound, Histopathology, Dermoscopy, Fundus, OCT, Digital Photography, Endoscopy, and PET.

Release

Technical report: Arxiv: Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning.
Model weights:
- Lingshu-7B
- Lingshu-32B

Disclaimer: We must note that even though the weights, codes, and demos are released in an open manner, similar to other pre-trained language models, and despite our best efforts in red teaming and safety fine-tuning and enforcement, our models come with potential risks, including but not limited to inaccurate, misleading or potentially harmful generation. Developers and stakeholders should perform their own red teaming and provide related security measures before deployment, and they must abide by and comply with local governance and regulations. In no event shall the authors be held liable for any claim, damages, or other liability arising from the use of the released weights, codes, or demos.

Evaluation

Medical Multimodal VQA

Models	MMMU-Med	VQA-RAD	SLAKE	PathVQA	PMC-VQA	OmniMedVQA	MedXpertQA	Avg.
Proprietary Models
GPT-4.1	75.2	65.0	72.2	55.5	55.2	75.5	45.2	63.4
Claude Sonnet 4	74.6	67.6	70.6	54.2	54.4	65.5	43.3	61.5
Gemini-2.5-Flash	76.9	68.5	75.8	55.4	55.4	71.0	52.8	65.1
Open-source Models (<10B)
BiomedGPT	24.9	16.6	13.6	11.3	27.6	27.9	-	-
Med-R1-2B	34.8	39.0	54.5	15.3	47.4	-	21.1	-
MedVLM-R1-2B	35.2	48.6	56.0	32.5	47.6	77.7	20.4	45.4
MedGemma-4B-IT	43.7	72.5	76.4	48.8	49.9	69.8	22.3	54.8
LLaVA-Med-7B	29.3	53.7	48.0	38.8	30.5	44.3	20.3	37.8
HuatuoGPT-V-7B	47.3	67.0	67.8	48.0	53.3	74.2	21.6	54.2
BioMediX2-8B	39.8	49.2	57.7	37.0	43.5	63.3	21.8	44.6
Qwen2.5VL-7B	50.6	64.5	67.2	44.1	51.9	63.6	22.3	52.0
InternVL2.5-8B	53.5	59.4	69.0	42.1	51.3	81.3	21.7	54.0
InternVL3-8B	59.2	65.4	72.8	48.6	53.8	79.1	22.4	57.3
Lingshu-7B	54.0	67.9	83.1	61.9	56.3	82.9	26.7	61.8
Open-source Models (>10B)
HealthGPT-14B	49.6	65.0	66.1	56.7	56.4	75.2	24.7	56.2
HuatuoGPT-V-34B	51.8	61.4	69.5	44.4	56.6	74.0	22.1	54.3
MedDr-40B	49.3	65.2	66.4	53.5	13.9	64.3	-	-
InternVL3-14B	63.1	66.3	72.8	48.0	54.1	78.9	23.1	58.0
Qwen2.5V-32B	59.6	71.8	71.2	41.9	54.5	68.2	25.2	56.1
InternVL2.5-38B	61.6	61.4	70.3	46.9	57.2	79.9	24.4	57.4
InternVL3-38B	65.2	65.4	72.7	51.0	56.6	79.8	25.2	59.4
Lingshu-32B	62.3	76.5	89.2	65.9	57.9	83.4	30.9	66.6

Medical Textual QA

Models	MMLU-Med	PubMedQA	MedMCQA	MedQA	Medbullets	MedXpertQA	SuperGPQA-Med	Avg.
Proprietary Models
GPT-4.1	89.6	75.6	77.7	89.1	77.0	30.9	49.9	70.0
Claude Sonnet 4	91.3	78.6	79.3	92.1	80.2	33.6	56.3	73.1
Gemini-2.5-Flash	84.2	73.8	73.6	91.2	77.6	35.6	53.3	69.9
Open-source Models (<10B)
Med-R1-2B	51.5	66.2	39.1	39.9	33.6	11.2	17.9	37.0
MedVLM-R1-2B	51.8	66.4	39.7	42.3	33.8	11.8	19.1	37.8
MedGemma-4B-IT	66.7	72.2	52.2	56.2	45.6	12.8	21.6	46.8
LLaVA-Med-7B	50.6	26.4	39.4	42.0	34.4	9.9	16.1	31.3
HuatuoGPT-V-7B	69.3	72.8	51.2	52.9	40.9	10.1	21.9	45.6
BioMediX2-8B	68.6	75.2	52.9	58.9	45.9	13.4	25.2	48.6
Qwen2.5VL-7B	73.4	76.4	52.6	57.3	42.1	12.8	26.3	48.7
InternVL2.5-8B	74.2	76.4	52.4	53.7	42.4	11.6	26.1	48.1
InternVL3-8B	77.5	75.4	57.7	62.1	48.5	13.1	31.2	52.2
Lingshu-7B	74.5	76.6	55.9	63.3	56.2	16.5	26.3	52.8
Open-source Models (>10B)
HealthGPT-14B	80.2	68.0	63.4	66.2	39.8	11.3	25.7	50.7
HuatuoGPT-V-34B	74.7	72.2	54.7	58.8	42.7	11.4	26.5	48.7
MedDr-40B	65.2	77.4	38.4	59.2	44.3	12.0	24.0	45.8
InternVL3-14B	81.7	77.2	62.0	70.1	49.5	14.1	37.9	56.1
Qwen2.5VL-32B	83.2	68.4	63.0	71.6	54.2	15.6	37.6	56.2
InternVL2.5-38B	84.6	74.2	65.9	74.4	55.0	14.7	39.9	58.4
InternVL3-38B	83.8	73.2	64.9	73.5	54.6	16.0	42.5	58.4
Lingshu-32B	84.7	77.8	66.1	74.7	65.4	22.7	41.1	61.8

Medical Report Generation

Models	MIMIC-CXR	CheXpert Plus	IU-Xray
ROUGE-L	CIDEr	RaTE	SembScore	RadCliQ-v1^-1	ROUGE-L	CIDEr	RaTE	SembScore	RadCliQ-v1^-1	ROUGE-L	CIDEr	RaTE	SembScore	RadCliQ-v1^-1
Proprietary Models
GPT-4.1	9.0	82.8	51.3	23.9	57.1	24.5	78.8	45.5	23.2	45.5	30.2	124.6	51.3	47.5	80.3
Claude Sonnet 4	20.0	56.6	45.6	19.7	53.4	22.0	59.5	43.5	18.9	43.3	25.4	88.3	55.4	41.0	72.1
Gemini-2.5-Flash	25.4	80.7	50.3	29.7	59.4	23.6	72.2	44.3	27.4	44.0	33.5	129.3	55.6	50.9	91.6
Open-source Models (<10B)
Med-R1-2B	19.3	35.4	40.6	14.8	42.4	18.6	37.1	38.5	17.8	37.6	16.1	38.3	41.4	12.5	43.6
MedVLM-R1-2B	20.3	40.1	41.6	14.2	48.3	20.9	43.5	38.9	15.5	40.9	22.7	61.1	46.1	22.7	54.3
MedGemma-4B-IT	25.6	81.0	52.4	29.2	62.9	27.1	79.0	47.2	29.3	46.6	30.8	103.6	57.0	46.8	86.7
LLaVA-Med-7B	15.0	43.4	12.8	18.3	52.9	18.4	45.5	38.8	23.5	44.0	18.8	68.2	40.9	16.0	58.1
HuatuoGPT-V-7B	23.4	69.5	48.9	20.0	48.2	21.3	64.7	44.2	19.3	39.4	29.6	104.3	52.9	40.7	63.6
BioMediX2-8B	20.0	52.8	44.4	17.7	53.0	18.1	47.9	40.8	21.6	43.3	19.6	58.8	40.1	11.6	53.8
Qwen2.5VL-7B	24.1	63.7	47.0	18.4	55.1	22.2	62.0	41.0	17.2	43.1	26.5	78.1	48.4	36.3	66.1
InternVL2.5-8B	23.2	61.8	47.0	21.0	56.2	20.6	58.5	43.1	19.7	42.7	24.8	75.4	51.1	36.7	67.0
InternVL3-8B	22.9	66.2	48.2	21.5	55.1	20.9	65.4	44.3	25.2	43.7	22.9	76.2	51.2	31.3	59.9
Lingshu-7B	30.8	109.4	52.1	30.0	69.2	26.5	79.0	45.4	26.8	47.3	41.2	180.7	57.6	48.4	108.1
Open-source Models (>10B)
HealthGPT-14B	21.4	64.7	48.4	16.5	52.7	20.6	66.2	44.4	22.7	42.6	22.9	81.9	50.8	16.6	56.9
HuatuoGPT-V-34B	23.5	68.5	48.5	23.0	47.1	22.5	62.8	42.9	22.1	39.7	28.2	108.3	54.4	42.2	59.3
MedDr-40B	15.7	62.3	45.2	12.2	47.0	24.1	66.1	44.7	24.2	44.7	19.4	62.9	40.3	7.3	48.9
InternVL3-14B	22.0	63.7	48.6	17.4	46.5	20.4	60.2	44.1	20.7	39.4	24.8	93.7	55.0	38.7	55.0
Qwen2.5VL-32B	15.7	50.2	47.5	17.1	45.2	15.2	54.8	43.4	18.5	40.3	18.9	73.3	51.3	38.1	54.0
InternVL2.5-38B	22.7	61.4	47.5	18.2	54.9	21.6	60.6	42.6	20.3	45.4	28.9	96.5	53.5	38.5	69.7
InternVL3-38B	22.8	64.6	47.9	18.1	47.2	20.5	62.7	43.8	20.2	39.4	25.5	90.7	53.5	33.1	55.2
Lingshu-32B	28.8	96.4	50.8	30.1	67.1	25.3	75.9	43.4	24.2	47.1	42.8	189.2	63.5	54.6	130.4

Usage

Using transformers

python

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info


# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "lingshu-medical-mllm/Lingshu-32B",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

processor = AutoProcessor.from_pretrained("lingshu-medical-mllm/Lingshu-32B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "example.png",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# Preparation for inference
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)

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
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)

Using vLLM

python

from vllm import LLM, SamplingParams
from qwen_vl_utils import process_vision_info
import PIL
from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("lingshu-medical-mllm/Lingshu-32B")
llm = LLM(model="lingshu-medical-mllm/Lingshu-32B", limit_mm_per_prompt = {"image": 4}, tensor_parallel_size=2, enforce_eager=True, trust_remote_code=True,)
sampling_params = SamplingParams(
            temperature=0.7,
            top_p=1,
            repetition_penalty=1,
            max_tokens=1024,
            stop_token_ids=[],
        )

text = "What does the image show?"
image_path = "example.png"
image = PIL.Image.open(image_path)

message = [
    {
        "role":"user",
        "content":[
            {"type":"image","image":image},
            {"type":"text","text":text}
            ]
            }
]
prompt = processor.apply_chat_template(
    message,
    tokenize=False,
    add_generation_prompt=True,
)
image_inputs, video_inputs = process_vision_info(message)
mm_data = {}
mm_data["image"] = image_inputs
processed_input = {
  "prompt": prompt,
  "multi_modal_data": mm_data,
}

outputs = llm.generate([processed_input], sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

Citation

If you find our project useful, we hope you would kindly star our repo and cite our work as follows:

* are equal contributions. ^ are corresponding authors.

@article{xu2025lingshu,
  title={Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning},
  author={Xu, Weiwen and Chan, Hou Pong and Li, Long and Aljunied, Mahani and Yuan, Ruifeng and Wang, Jianyu and Xiao, Chenghao and Chen, Guizhen and Liu, Chaoqun and Li, Zhaodonghui and others},
  journal={arXiv preprint arXiv:2506.07044},
  year={2025}
}

🚀 If you find these models useful

Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:

👉 Quantum Network Monitor

The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder

💬 How to test:
Choose an AI assistant type:

TurboLLM (GPT-4.1-mini)
HugLLM (Hugginface Open-source models)
TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

Function calling against live network services
How small can a model go while still handling:
- Automated Nmap security scans
- Quantum-readiness checks
- Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):

✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4.1-mini :

**It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
Create custom cmd processors to run .net code on Quantum Network Monitor Agents
Real-time network diagnostics and monitoring
Security Audits
Penetration testing (Nmap/Metasploit)

🔵 HugLLM – Latest Open-source models:

🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.

💡 Example commands you could test:

"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!

Final Word

I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.

If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.

I'm also open to job opportunities or sponsorship.

Thank you! 😊

Mungert/Lingshu-32B-GGUF

Lingshu-32B GGUF Models

Model Generation Details

Quantization Beyond the IMatrix