---
title: EvoQwen2.5-VL-Retriever-7B-v1
canonical_url: "https://www.modelscope.cn/models/Apsara-Stack-MaaS/EvoQwen2.5-VL-Retriever-7B-v1"
md_url: "https://www.modelscope.cn/models/Apsara-Stack-MaaS/EvoQwen2.5-VL-Retriever-7B-v1.md"
repository: Apsara-Stack-MaaS/EvoQwen2.5-VL-Retriever-7B-v1
last_updated: 2026-04-09
pipeline_tag: image-text-retrieval
tasks:
- image-text-retrieval
model_type:
- qwen2_5_vl
architectures:
- ColQwen2_5
parameters: 7.7B
tensor_type:
- BF16
library_name:
- safetensors
- pytorch
frameworks:
- Pytorch
downloads: 2052
stars: 9
---
# EvoQwen2.5-VL-Retriever-7B-v1
> EvoQwen2.5-VL-Retriever-7B-v1 - Apsara-Stack-MaaS 在 ModelScope 开源的模型。EvoQwen2.5-VL-Retriever-7B-v1 是一个基于 Qwen2.5-VL-7B-Instruct 主干网络和多向量后期交互（multi-vector late-interaction）的高性能多模态检索模型。该模型通过创新性的进化式训练框架（Evo-Retriever）微调，从而实现精准的复杂视觉文档检索能力。
- **Repository**: Apsara-Stack-MaaS/EvoQwen2.5-VL-Retriever-7B-v1
- **Tasks**: image-text-retrieval
- **Parameters**: 7.7B
- **Downloads**: 2052
- **Stars**: 9
- **Last updated**: 2026-04-09
Source: https://www.modelscope.cn/models/Apsara-Stack-MaaS/EvoQwen2.5-VL-Retriever-7B-v1
---
EvoQwen2.5-VL-Retriever-7B-v1 是一个基于 Qwen2.5-VL-7B-Instruct 主干网络和多向量后期交互（multi-vector late-interaction）的高性能多模态检索模型。该模型通过创新性的进化式训练框架（Evo-Retriever）微调，从而实现精准的复杂视觉文档检索能力。
**版本特性**** ****(Version Specificity)**
+ **基础模型**** ****(Base Model)**: ApsaraStackMaaS/EvoQwen2.5-VL-Retriever-7B-v1
+ **参数规模**** ****(Size)**: 70亿 (7B)
+ **特点**** ****(Features)**: 7B版本在所有评估基准上均提供最高的检索精度。它非常适合对性能要求极致的应用场景。
**性能**** ****(Performance)**
| **模型** | ViDoRe V2（nDCG@5） | MMEB VisDoc（ndcg_linear@5） |
| :--- | :--- | :--- |
| ApsaraStackMaaS/EvoQwen2.5-VL-Retriever-3B-v1 | **63.00** | **75.96** |
| ApsaraStackMaaS/EvoQwen2.5-VL-Retriever-7B-v1 | 65.24 | **77.10** |
**使用方法**** ****(Usage)**
请确保已安装 transformers, torch, Pillow,colpali_engine 。
```python
import torch
from PIL import Image
from transformers.utils.import_utils import is_flash_attn_2_available
from colpali_engine.models import ColQwen2_5, ColQwen2_5_Processor
model_name = "ApsaraStackMaaS/EvoQwen2.5-VL-Retriever-7B-v1"
model = ColQwen2_5.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="cuda:0", # or "mps" if on Apple Silicon
attn_implementation="flash_attention_2" if is_flash_attn_2_available() else None,
).eval()
processor = ColQwen2_5_Processor.from_pretrained(model_name)
# Your inputs
images = [
Image.new("RGB", (128, 128), color="white"),
Image.new("RGB", (64, 32), color="black"),
]
queries = [
"Is attention really all you need?",
"What is the amount of bananas farmed in Salvador?",
]
# Process the inputs
batch_images = processor.process_images(images).to(model.device)
batch_queries = processor.process_queries(queries).to(model.device)
# Forward pass
with torch.no_grad():
image_embeddings = model(**batch_images)
query_embeddings = model(**batch_queries)
scores = processor.score_multi_vector(query_embeddings, image_embeddings)
print(scores)
```
**Parameters / 参数**
所有模型都使用Evo-Retriever 训练范式进行一个为期两阶段、每阶段一个epoch的训练课程进行微调。除非另有说明，我们采用低秩适配器（LoRA）进行参数高效的微调，其中3B和7B模型的LoRA秩（rank）均设置为32。我们以 bfloat16 格式训练模型，并使用 paged_adamw_8bit 优化器。训练在一个包含8块H20 GPU的服务器上进行，采用数据并行策略，学习率为 2e-5 并配合余弦衰减和2%的预热步数，批量大小为32。
**Citation / 引用**
```python
@misc{li2026evoretriever,
title={Evo-Retriever: LLM-Guided Curriculum Evolution with Viewpoint-Pathway Collaboration for Multimodal Document Retrieval},
author={Weiqing Li and Jinyue Guo and Yaqi Wang and Haiyang Xiao and Yuewei Zhang and Guohua Liu and Hao Henry Wang},
year={2026},
eprint={2603.16455},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.16455},
}
```