---
title: siglip2-base-p32-256-ve
canonical_url: "https://www.modelscope.cn/models/gongjy/siglip2-base-p32-256-ve"
md_url: "https://www.modelscope.cn/models/gongjy/siglip2-base-p32-256-ve.md"
repository: gongjy/siglip2-base-p32-256-ve
chinese_name: siglip2-base-p32-256-ve
last_updated: 2026-04-20
license: "Apache License 2.0"
model_type:
  - siglip_vision_model
architectures:
  - SiglipVisionModel
base_model:
  - google/siglip2-base-patch32-256
base_model_relation: repackage
parameters: 94.6M
tensor_type:
  - F16
library_name:
  - transformer
  - safetensors
language:
  - zh
  - en
downloads: 4308
stars: 1
---

# siglip2-base-p32-256-ve

> siglip2-base-p32-256-ve - gongjy 在 ModelScope 开源的模型。SigLIP2 ViT-B/32 Vision Encoder (256)

gongjy/siglip2-base-p32-256-ve 是 ModelScope 魔搭社区上的 94.6M 参数机器学习模型，采用 Apache License 2.0 许可，基于 google/siglip2-base-patch32-256 构建。

- **Repository**: gongjy/siglip2-base-p32-256-ve
- **License**: Apache License 2.0
- **Parameters**: 94.6M
- **Base model**: google/siglip2-base-patch32-256
- **Downloads**: 4308
- **Stars**: 1
- **Last updated**: 2026-04-20

Source: https://www.modelscope.cn/models/gongjy/siglip2-base-p32-256-ve

---

# SigLIP2 ViT-B/32 Vision Encoder (256)

模型来源：[google/siglip2-base-patch32-256](https://huggingface.co/google/siglip2-base-patch32-256)

仅裁剪保留了 **Vision Encoder** 部分，去除了 Text Encoder 及其他无关权重，并以 **float16** 精度保存为 HuggingFace `SiglipVisionModel` 兼容的 safetensors 格式。

---

Source: [google/siglip2-base-patch32-256](https://huggingface.co/google/siglip2-base-patch32-256)

Only the **Vision Encoder** is retained. The Text Encoder and all unrelated weights have been removed. Weights are saved in **float16** precision as HuggingFace `SiglipVisionModel`-compatible safetensors format.

---
## Model Info

| Item | Value |
|------|-------|
| Architecture | SigLIP2 ViT-B/32 (fixed resolution) |
| Parameters | 94.6M |
| Hidden Size | 768 |
| Layers | 12 |
| Patch Size | 32 |
| Input Resolution | 256×256 |
| Output Tokens | 64 (8×8) |
| Weight Precision | float16 |
| File Size | ~180MB |

## Usage

```python
from transformers import SiglipVisionModel, SiglipImageProcessor

model = SiglipVisionModel.from_pretrained("path_to/siglip2-base-p32-256-ve")
processor = SiglipImageProcessor.from_pretrained("path_to/siglip2-base-p32-256-ve")

inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)  # outputs.last_hidden_state: [B, 64, 768]
```