---
title: LLaVA-OneVision-Qwen2.5-72B-Instruct
canonical_url: "https://www.modelscope.ai/models/wangq2k/LLaVA-OneVision-Qwen2.5-72B-Instruct"
md_url: "https://www.modelscope.ai/models/wangq2k/LLaVA-OneVision-Qwen2.5-72B-Instruct.md"
repository: wangq2k/LLaVA-OneVision-Qwen2.5-72B-Instruct
last_updated: 2026-10-24
parameters: 292.7B
tensor_type:
  - BF16
library_name:
  - safetensors
downloads: 9
stars: 0
---

# LLaVA-OneVision-Qwen2.5-72B-Instruct

> LLaVA-OneVision-Qwen2.5-72B-Instruct - An open-source model by wangq2k on ModelScope. This model is a vision-language model (VLM) trained based on the LLaVA-OneVision framework. The Qwen2.5-72B-Instruct model is adopted as the language model backbone, while…

- **Repository**: wangq2k/LLaVA-OneVision-Qwen2.5-72B-Instruct
- **Parameters**: 292.7B
- **Downloads**: 9
- **Stars**: 0
- **Last updated**: 2026-10-24

Source: https://www.modelscope.ai/models/wangq2k/LLaVA-OneVision-Qwen2.5-72B-Instruct

---

# Model Card

## Introduction

This model is a vision-language model (VLM) trained based on the [LLaVA-OneVision](https://llava-vl.github.io/blog/2024-08-05-llava-onevision/) framework. The [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) model is adopted as the language model backbone, while [SigLIP-SO400M-Patch14-384](https://huggingface.co/google/siglip-so400m-patch14-384) is used as the vision encoder.

The model uses a two-layer MLP projector to map visual features into the language model's embedding space.

## Training Data

The model is trained on a subset of the [Infinity-MM](https://huggingface.co/datasets/BAAI/Infinity-MM) dataset. Due to computational constraints, selected portions of the dataset are used rather than the full corpus.

- **Stage 1 (visual-language alignment):** Approximately 5M samples are drawn from a 50% random subset of the Stage 1 partition. Only the projector is trained.
- **Stage 2 (multimodal fine-tuning):** A combined dataset of approximately 12M samples is constructed from the complete Stages 3–4 data and approximately 3M samples from the Stage 2 partition. The full model is fine-tuned end-to-end.

## How to Use

Inference requires a GPU and the model adaptations provided in our [code repository](https://github.com/wangq-dev/CDMScaling-Inference). See the repository's `README.md` for environment setup, installation, and instructions for applying the model-specific changes to LLaVA-NeXT.

For single-image inference with a trained checkpoint, use the provided `infer_llava_gpu.py` script:

```bash
python infer_llava_gpu.py \
    --model-path /path/to/model/checkpoint \
    --model-name llava_qwen \
    --image /path/to/image.jpg \
    --conv-template qwen_2_5 \
    --question "What is shown in this image?" \
    --device-map auto \
    --max-new-tokens 512
```

Replace the checkpoint and image paths with your own paths. The `--device-map auto` option enables automatic device placement.

## Citation

If you find this work useful, please cite the following paper:

```bibtex
@misc{li2026transferstextvisioncapability,
    title={What Transfers from Text to Vision? Capability Scaling Laws and Transfer Dynamics for VLMs},
    author={Ziran Li and Qiang Wang and Zhengyu Chen and Shanglin Lei and Borun Chen and Jingang Wang and Xunliang Cai},
    year={2026},
    eprint={2608.00013},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2608.00013},
}
```