README.md

---
license: apache-2.0
datasets:
- AIDC-AI/Ovis-dataset
library_name: transformers
tags:
- MLLM
- ovis
- qwen3
pipeline_tag: image-text-to-text
language:
- en
- vi
- zh
---

# Ovis2.5-2B-Pretrained (Qwen3-1.7B + SigLIP2) - Final Version For Pretraining

<div align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/637aebed7ce76c3b834cea37/3IK823BZ8w-mz_QfeYkDn.png" width="30%"/>
</div>

<p align="center">
  <a href="https://arxiv.org/abs/2508.11737"><img src="https://img.shields.io/badge/📖_Original_Report-Ovis2.5-b31b1b.svg" alt="technical report"></a>
  <a href="https://github.com/AIDC-AI/Ovis"><img src="https://img.shields.io/badge/GitHub-AIDC--AI/Ovis-blue?style=flat&logo=github" alt="code"></a>
  <a href="https://huggingface.co/collections/AIDC-AI/ovis25-689ec1474633b2aab8809335"><img src="https://img.shields.io/badge/🤗_Official_Models-AIDC--AI/Ovis2.5-yellow" alt="models"></a>
</p>

---

# Ovis2.5-2B-Pretrained (Qwen3-1.7B + SigLIP2)

**Ovis2.5-2B-Pretrained** is a merged version combining:

- **Vision Encoder:** `siglip2-so400m-patch16-512` (from Ovis2.5)
- **Language Model (LLM):** `Qwen3-1.7B` (lightweight, efficient, supports Vietnamese)

> **Note:** This is a base/pretrained model, only merged weights, not instruction-tuned. For best conversational performance, further fine-tuning is required.

## Architecture Details

| Ovis MLLM                | Vision Encoder                | Language Model (LLM) | Status                        |
|--------------------------|-------------------------------|----------------------|-------------------------------|
| VOvis2.5-2B-Pretrained(Final Version)    | siglip2-so400m-patch16-512    | Qwen3-1.7B           | Base PT Model (Needs SFT)|
| Ovis2.5-2B (Official)    | siglip2-so400m-patch16-512    | Qwen3-1.7B           | Instruction-Tuned             |
| Ovis2.5-9B (Official)    | siglip2-so400m-patch16-512    | Qwen3-8B             | Instruction-Tuned             |

**Supported languages:** Vietnamese 🇻🇳, English, Chinese
---

## 🚀 **Quick Start**

### **Cài đặt**
```bash
pip install torch==2.8.0 transformers==4.51.3 numpy==1.26.4
pip install flash-attn==2.7.4.post1 --no-build-isolation
```

## Quick Start

```python
import torch
from PIL import Image
from transformers import AutoModelForCausalLM
import requests

model = AutoModelForCausalLM.from_pretrained(
    "AIDC-AI/VOvis2.5-2B-pt",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
).cuda()

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": Image.open(requests.get("https://cdn-uploads.huggingface.co/production/uploads/658a8a837959448ef5500ce5/TIlymOb86R6_Mez3bpmcB.png", stream=True).raw)},
        {"type": "text", "text": "Describe the image in detail."},
    ],
}]

input_ids, pixel_values, grid_thws = model.preprocess_inputs(
    messages=messages,
    add_generation_prompt=True,
    enable_thinking=True
)
input_ids = input_ids.cuda()
pixel_values = pixel_values.cuda() if pixel_values is not None else None
grid_thws = grid_thws.cuda() if grid_thws is not None else None

outputs = model.generate(
    inputs=input_ids,
    pixel_values=pixel_values,
    grid_thws=grid_thws,
    enable_thinking=True,
    enable_thinking_budget=True,
    max_new_tokens=3072,
    thinking_budget=1024,
)

response = model.text_tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```