README.md

---
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
datasets:
- HuggingFaceM4/FineVision
- mvp-lab/LLaVA-OneVision-1.5-Instruct-Data
language:
- en
license: cc-by-nc-sa-4.0
pipeline_tag: image-text-to-text
library_name: transformers
---

# CASA-Qwen2_5-VL-3B

This repository contains the model weights for **CASA-Qwen2_5-VL-3B**, introduced in the paper [CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion](https://huggingface.co/papers/2512.19535).

This model is a [Qwen-2.5VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) model adapted from token insertion to a cross-attention-based architecture.

- **Paper:** [CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion](https://arxiv.org/abs/2512.19535)
- **Project Page:** [kyutai.org/casa](https://kyutai.org/casa)
- **Code:** [github.com/kyutai-labs/casa](https://github.com/kyutai-labs/casa)

## Sample Usage

This model requires `trust_remote_code=True` to load the custom architecture. Below is a snippet to run inference using `transformers`.

```python
import torch
from transformers.models.auto.modeling_auto import AutoModel
from transformers.models.auto.processing_auto import AutoProcessor

model_id = "kyutai/CASA-Qwen2_5-VL-3B"
model = AutoModel.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    trust_remote_code=True,
).cuda()

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
)

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/ai2d-demo.png",
            },
            {
                "type": "text",
                "text": "Describe this image.",
            },
        ],
    },
]

inputs = processor.tokenize_messages(messages=conversation)
inputs = inputs.to(model.device)
input_len = inputs["input_ids"].shape[1]

output_ids = model.generate_from_image(
  **inputs,
  max_new_tokens=512,
  pre_image_tokens=processor.pre_image_tokens,
  post_image_tokens=processor.post_image_tokens,
  eos_token_id=model.generation_config.eos_token_id,
)[0, input_len:]

response = processor.tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)
```

## Citation

```bibtex
@article{kyutai2025casa,
  author = {Moritz B\"ohle and Am\'elie Royer and Juliette Marrie and Edouard Grave and Patrick P\'erez},
  year = {2025},
  title = {CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion},
  journal = {ArXiv},
  url = {https://arxiv.org/abs/2512.19535}
}
```

## License

The code in the official repository is provided under the **MIT license**. The weights for this model are released under the **CC-BY-NC-SA 4.0 license**. Additionally, as this model includes weights from Qwen2.5-VL-3B, it is subject to the [Qwen RESEARCH LICENSE AGREEMENT](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/blob/main/LICENSE).