Instructions to use yongxianwei/Qwen2.5-VL-32B-OCR with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use yongxianwei/Qwen2.5-VL-32B-OCR with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="yongxianwei/Qwen2.5-VL-32B-OCR")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("yongxianwei/Qwen2.5-VL-32B-OCR")
model = AutoModelForMultimodalLM.from_pretrained("yongxianwei/Qwen2.5-VL-32B-OCR", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use yongxianwei/Qwen2.5-VL-32B-OCR with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "yongxianwei/Qwen2.5-VL-32B-OCR"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yongxianwei/Qwen2.5-VL-32B-OCR",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/yongxianwei/Qwen2.5-VL-32B-OCR

SGLang

How to use yongxianwei/Qwen2.5-VL-32B-OCR with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "yongxianwei/Qwen2.5-VL-32B-OCR" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yongxianwei/Qwen2.5-VL-32B-OCR",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "yongxianwei/Qwen2.5-VL-32B-OCR" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yongxianwei/Qwen2.5-VL-32B-OCR",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use yongxianwei/Qwen2.5-VL-32B-OCR with Docker Model Runner:
```
docker model run hf.co/yongxianwei/Qwen2.5-VL-32B-OCR
```

Qwen2.5-VL-32B-OCR

The OCR expert of the OptMerge MLLM merging benchmark (ICLR 2026). It is one of five capability experts (VQA, Geometry, Chart, OCR, Grounding) fine-tuned from the same base model, so that they can be merged into one model without training data.

Paper: OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging (OpenReview)
Code: WalkerWorldPeace/MLLMerging
All experts: MLLMerging collection

Training

Base model: Qwen/Qwen2.5-VL-32B-Instruct
Method: Supervised fine-tuning of Qwen2.5-VL-32B-Instruct (see "Model scales" in the paper). This repository stores full weights; load it directly, no PEFT adapter is needed.
Data: OCRVQA, TextCaps, SynthDoG, LLaVAR, ST-VQA, TextVQA, DocVQA, DeepForm, KLC, TabFact (238K samples, Table 1 of the paper), plus additional high-quality fine-tuning data.

Results

From Table 9 of the paper. The last row merges this expert with the other four experts using OptMerge.

Model	VizWiz	GQA	MathVista	MATH-Vision	ChartQA	TextVQA	OCRVQA	RefCOCO	RefCOCO+	RefCOCOg	Avg.
Qwen2.5-VL-32B-Instruct (base model)	41.39	59.34	79.21	47.36	83.64	79.62	64.58	88.01	82.41	84.06	70.96
Qwen2.5-VL-32B-OCR	42.65	61.04	75.28	34.21	87.00	81.42	67.32	89.63	83.76	84.62	70.69
OptMerge (5 experts merged)	43.52	62.50	80.01	43.42	88.92	81.91	66.37	89.94	83.97	84.68	72.52

Usage

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "yongxianwei/Qwen2.5-VL-32B-OCR"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
# Build chat messages and call model.generate(...) as in the official Qwen2.5-VL model card.

To merge the five experts yourself, see Reproducing OptMerge.

Citation

If you use these checkpoints or the MLLMerging code, please cite both OptMerge and its extended version:

@inproceedings{wei2025optmerge,
  title={OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging},
  author={Wei, Yongxian and Cheng, Runxi and Jin, Weike and Yang, Enneng and Shen, Li and Hou, Lu and Du, Sinan and Yuan, Chun and Cao, Xiaochun and Tao, Dacheng},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=Me0n0iESJY}
}

@article{wei2026swudi,
  title={Closed-Form Spectral Regularization for Multi-Task Model Merging},
  author={Wei, Yongxian and Cheng, Runxi and Zhang, Xingxuan and Shen, Li and Yuan, Chun and Cui, Peng and Tao, Dacheng},
  journal={arXiv preprint arXiv:2606.07289},
  year={2026}
}