---
license: apache-2.0
library_name: transformers
pipeline_tag: video-text-to-text
---

# OmniVideo-30B (Qwen3-Omni)

[![Project Page](https://img.shields.io/badge/Project-Page-orange)](https://yzlmhzz.github.io/OmniVideo-100K/)
[![Paper](https://img.shields.io/badge/arXiv-Paper-b31b1b.svg)](https://arxiv.org/abs/2606.14702)
[![Github](https://img.shields.io/badge/Github-OmniVideo--100K-blue)](https://github.com/MiG-NJU/OmniVideo-100K)
[![Dataset](https://img.shields.io/badge/Dataset-OmniVideo--100K-green)](https://huggingface.co/datasets/MiG-NJU/OmniVideo-100K)

This is the fine-tuned **OmniVideo-30B** model, initialized from the official **Qwen3-Omni-30B-A3B-Instruct** and trained on the **[OmniVideo-100K](https://huggingface.co/datasets/MiG-NJU/OmniVideo-100K)** instruction-tuning dataset introduced in our paper: *"[OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains](https://arxiv.org/abs/2606.14702)"*.

## 🚀 Performance

<div style="overflow-x: auto;">
  <table style="white-space: nowrap; width: 100%; text-align: left;">
    <thead>
      <tr>
        <th style="text-align: left;">Models</th>
        <th>OmniVideo-Test</th>
        <th>Daily-Omni</th>
        <th>OmniVideoBench</th>
        <th>JointAVBench</th>
        <th>FutureOmni</th>
        <th>Video-MME<sub>short</sub></th>
        <th>Video-MME-v2</th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td style="text-align: left;">Qwen3-Omni-30B-A3B-Instruct</td>
        <td>49.70</td>
        <td>74.27</td>
        <td>43.84</td>
        <td>63.17</td>
        <td>53.44</td>
        <td>82.00</td>
        <td>14.31</td>
      </tr>
      <tr>
        <td style="text-align: left;">OmniVideo-30B <sub>(Qwen3-Omni)</sub></td>
        <td><b>63.56</b><sub><font color="red">+13.86</font></sub></td>
        <td><b>76.61</b><sub><font color="red">+2.34</font></sub></td>
        <td><b>44.81</b><sub><font color="red">+0.97</font></sub></td>
        <td><b>66.37</b><sub><font color="red">+3.20</font></sub></td>
        <td><b>57.60</b><sub><font color="red">+4.16</font></sub></td>
        <td><b>83.56</b><sub><font color="red">+1.56</font></sub></td>
        <td><b>15.33</b><sub><font color="red">+1.02</font></sub></td>
      </tr>
    </tbody>
  </table>
</div>


## ⚙️ Training and Evaluation

The model was fine-tuned full-parameter using `LLaMA-Factory`.

For detailed training and evaluation instructions, please refer to the official [OmniVideo-100K GitHub repository](https://github.com/MiG-NJU/OmniVideo-100K). 

## 📑 Citation

If you find this work useful, please cite our paper:

```bibtex
@article{cai2026omnivideo100k,
  title={OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains},
  author={Cai, Xinyue suicide and Fu, Chaoyou and Zhang, Yi-Fan and He, Ran and Shan, Caifeng},
  journal={arXiv preprint arXiv:2606.14702}, 
  year={2026}
}
```