README.md

---
license: apache-2.0
datasets:
- lmms-lab/LLaVA-OneVision-Data
- BAAI/Infinity-MM
language:
- en
- zh
base_model:
- google/siglip2-so400m-patch16-512
- Qwen/Qwen2-1.5B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
---

# Flash-VL-2B-Static
[\[📜 Flash-VL Tech Report\]](https://www.arxiv.org/abs/2505.09498)

![image/png](https://s3plus.meituan.net/automl-datasets/mlm/logo.jpg)

## Introduction

We are excited to introduce **Flash-VL**, a novel approach to optimizing Vision-Language Models (VLMs) for real-time applications, targeting ultra-low latency and high throughput without sacrificing accuracy. Leveraging advanced architectural enhancements and efficient computational strategies, Flash-VL 2B is designed to maximize throughput by reducing processing time while maintaining competitive performance across multiple vision-language benchmarks. Our approach includes tailored architectural choices, token compression mechanisms, data curation, training schemes, and a novel image processing technique called implicit semantic stitching that effectively balances computational load and model performance. Through extensive evaluations on 11 standard VLM benchmarks, we demonstrate that Flash-VL 2B achieves state-of-the-art results in both speed and accuracy, making it a promising solution for deployment in resource-constrained environments and large-scale real-time applications.


### Environment Setup

```bash
pip install torch==2.1.2
pip install transformers==4.50.0.dev0
```


### How to use it?

```python
import torch
from PIL import Image
import requests
from io import BytesIO
from transformers import AutoModel, AutoTokenizer, AutoProcessor

model_path = "FlashVL/FlashVL-2B-Static"
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16,trust_remote_code=True,device_map='cuda')
model.tokenizer = AutoTokenizer.from_pretrained(model_path,device_map='cuda')
model.im_trans = AutoProcessor.from_pretrained(model_path).image_processor

# single-image single-round conversation (单图单轮对话)
image_url ="https://s3plus.meituan.net/automl-datasets/mlm/0516.png"
response = requests.get(image_url)
image_data = BytesIO(response.content)
pil_image = Image.open(image_data).convert('RGB')   
messages = [{'role': 'user', 'content': "生成图中菜品的菜谱"}] # answer: EXTRA
answer = model.chat(pil_image, messages, do_sample=False, max_new_tokens=256)
print(answer)

# single-image multi-round conversation (单图多轮对话)
messages = [
    {'role': 'user', 'content': '这是什么'},
    {"role": "assistant", "content": '这是一道看起来像是银耳莲子汤的甜品。\
     银耳是一种常见的食材，通常用于制作甜品和汤品，具有软糯的口感和清润的口感。莲 \
     子是莲子的干燥部分，常用于中医和食疗中，具有补脾止泻的功效。图片中还可以看到 \
     一些枸杞和核桃，枸杞富含维生素和抗氧化物质，核桃则提供丰富的蛋白质和健康脂肪。 \
     整体来看，这道甜品不仅美味，还具有一定的营养价值。'},
    {'role': 'user', 'content': '对图中菜品卡路里分析'}
    ]
answer = model.chat(pil_image, messages, do_sample=False, max_new_tokens=256)
print(answer)

# pure-text single-round conversation (纯文本对话）
messages = [{'role': 'user', 'content': "who are you"}]
answer = model.chat(None, messages, do_sample=False, max_new_tokens=256)
print(answer)
```

### Evaluation

| Benchmark | Qwen2-VL-2B | Aquila-VL-2B | InternVL2.5-2B | Flash-VL-2B<sub>s<sub> | Flash-VL-2B<sub>d<sub> | Flash-VL-2B<sub>d-ISS<sub> |
| :-------------: | :-------------: | :-------------: | :-------------: |:-------------: |:-------------: |:-------------: |
| MMMU<sub>val<sub>                | 41.9 | 44.4 | 41.8 | 43.6 | 42.9 | 42.9 |
| MMBench<sup>en<sup>              | 74.9 | 78.6 | 74.7 | 78.4 | 78.4 | 79.1 |
| MMBench<sup>cn<sup>              | 73.5 | 76.3 | 71.6 | 74.7 | 74.9 | 76.7 |
| MMStar                           | 48.0 | 54.9 | 54.1 | 53.8 | 54.4 | 54.1 |
| MathVista<sub>testmini<sub>      | 43.0 | 59.4 | 50.9 | 59.3 | 58.1 | 61.5 |
| AI2D<sub>test<sub>               | 74.1 | 75.0 | 75.1 | 74.2 | 74.1 | 74.4 |
| MMVet                            | 49.5 | 40.9 | 61.7 | 47.3 | 52.7 | 50.7 |
| HallusionBench                   | 39.2 | 38.5 | 42.7 | 43.5 | 45.5 | 49.0 |
| OCRBench                         |  794 |  773 | 800  | 764  | 831  | 843  |
| MME                              | 1872 | 1813 | 2091 | 1715 | 1866 | 1850 |
| SEEDBench                        | 71.5 | 78.9 | 73.2 | 73.6 | 73.6 | 74.5 |
| Average                          | 60.2 | 62.6 | 63.6 | 62.4 | 64.0 | 64.8 |


We use [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) to evaluate FlashVL-2B-Static. 



## Citation
If you find this project useful in your research, please consider citing:

```BibTeX
@misc{zhang2025flashvl2boptimizingvisionlanguage,
      title={Flash-VL 2B: Optimizing Vision-Language Model Performance for Ultra-Low Latency and High Throughput}, 
      author={Bo Zhang and Shuo Li and Runhe Tian and Yang Yang and Jixin Tang and Jinhao Zhou and Lin Ma},
      year={2025},
      eprint={2505.09498},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.09498}, 
}
```