---
license: cc-by-nc-4.0
base_model: Qwen/Qwen3.5-9B
base_model_relation: finetune
datasets:
- mxlin043/EmbRACE
language:
- en
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- embodied-ai
- vision-language-navigation
- egocentric
---

# EmbRACE-Qwen3.5-9B

[Dataset](https://huggingface.co/datasets/mxlin043/EmbRACE) · [Viewer](https://huggingface.co/spaces/mxlin043/EmbRACE-Viewer) · [Paper](https://arxiv.org/abs/2507.10548) · [Code](https://github.com/mxlin043/EmbRACE)

[Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) fine-tuned on the dataset part of EmbRACE. The model acts as an embodied agent in the closed loop. At every step it receives the instruction and its egocentric frames and answers with a rationale and the next action.

## Interface

The model is run with the protocol of the EmbRACE evaluation code, which builds the prompts. Each task is one multi-turn conversation whose system prompt holds the instruction and the action space. The conversation keeps the initial frame and the 13 most recent frames, the last of which is the current one, so a request carries at most 14 images. Every earlier frame is followed by the model's reply to it. The model answers

```
<reasoning>...</reasoning><action>ACTION</action>
```

where `ACTION` is one of `MoveForward`, `MoveBackward`, `TurnLeft`, `TurnRight`, `LookUp`, `LookDown`, `OpenDoor`, `Pick`, `Drop`, `MidwayTarget` and `Finish`.

## Usage

Serve the model with vLLM.

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve mxlin043/EmbRACE-Qwen3.5-9B \
    --served-model-name embrace-qwen35-9b \
    --host 127.0.0.1 --port 8000 \
    --dtype bfloat16 \
    --max-model-len 32768 \
    --limit-mm-per-prompt '{"image": 16}'
```

Send the following fields with every request, as the evaluation code does. The checkpoint does not set them as defaults.

| Field | Value |
|---|---|
| `temperature` | 0.7 |
| `top_p` | 0.8 |
| `top_k` | 20 |
| `min_p` | 0.0 |
| `presence_penalty` | 1.5 |
| `chat_template_kwargs` | `{"enable_thinking": false}` |

To run the benchmark with the [evaluation code](https://github.com/mxlin043/EmbRACE), add an entry to `configs/vlm_models.yaml`

```yaml
  - match: embrace-qwen35-9b
    provider: local_vllm
    sampling: qwen_instruct
    extra_body:
      chat_template_kwargs: {enable_thinking: false}
```

and start the evaluation with the simulator on another GPU. The README of the code describes how to install the simulator and download the benchmark.

## License

The weights are released under the [Creative Commons Attribution-NonCommercial 4.0](LICENSE) license, the license of the EmbRACE data, for research use with attribution. The base model Qwen3.5-9B is released by the Qwen team under the Apache 2.0 license, a copy of which is included as [LICENSE-Qwen3.5-9B](LICENSE-Qwen3.5-9B).

## Citation

```bibtex
@article{lin2025embrace,
  title={EmbRACE: Embodied Reasoning and Action in Complex Environments},
  author={Lin, Mingxian and Huang, Wei and Li, Yitang and Jiang, Chengjie and Wu, Kui and Zhong, Fangwei and Chen, Weikai and Qian, Shengju and Wang, Xin and Qi, Xiaojuan},
  journal={arXiv preprint arXiv:2507.10548},
  year={2025}
}
```