---
title: QwenPaw-Flash-9B
canonical_url: "https://www.modelscope.cn/models/AgentScope/QwenPaw-Flash-9B"
md_url: "https://www.modelscope.cn/models/AgentScope/QwenPaw-Flash-9B.md"
repository: AgentScope/QwenPaw-Flash-9B
last_updated: 2026-11-02
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model:
  - Qwen/Qwen3.5-9B
base_model_relation: finetune
parameters: 9.4B
tensor_type:
  - BF16
library_name:
  - safetensors
downloads: 12
stars: 0
tags:
  - qwenpaw
  - tool-use
---

# QwenPaw-Flash-9B

> QwenPaw-Flash-9B - AgentScope 在 ModelScope 开源的模型。Mirror of agentscope-ai/QwenPaw-Flash-9B

AgentScope/QwenPaw-Flash-9B 是 ModelScope 魔搭社区上的 9.4B 参数text-generation模型，采用 apache-2.0 许可，基于 Qwen/Qwen3.5-9B 构建。

- **Repository**: AgentScope/QwenPaw-Flash-9B
- **License**: apache-2.0
- **Tasks**: text-generation
- **Parameters**: 9.4B
- **Base model**: Qwen/Qwen3.5-9B
- **Tags**: qwenpaw, tool-use
- **Downloads**: 12
- **Stars**: 0
- **Last updated**: 2026-11-02

Source: https://www.modelscope.cn/models/AgentScope/QwenPaw-Flash-9B

---

# QwenPaw-Flash-9B

QwenPaw-Flash-9B is the Safetensors release of the QwenPaw-Flash-9B agent model. The QwenPaw-Flash family is trained for tool use, command execution, memory management and multi-step tasks in QwenPaw.

[Project](https://github.com/agentscope-ai/QwenPaw) · [Hugging Face](https://huggingface.co/agentscope-ai/QwenPaw-Flash-9B) · [ModelScope collection](https://www.modelscope.cn/collections/AgentScope/QwenPaw-Flash)

## Download from ModelScope

With the ModelScope CLI installed, download this repository into one directory.

```bash
modelscope download --model AgentScope/QwenPaw-Flash-9B --local_dir QwenPaw-Flash-9B
```

## Weight files

The weight files are listed below.

| File | Bytes | SHA-256 |
| --- | ---: | --- |
| `model.safetensors` | 18,819,722,392 | `d55ed250d8bac00f677ce35c653c06d3abc079a26cd9bd2da2740ae8faf1399b` |

## Serve locally

Start a vLLM server with the downloaded checkpoint. For tool use, enable the Qwen tool-call parser.

```bash
vllm serve ./QwenPaw-Flash-9B --served-model-name QwenPaw-Flash-9B --port 8000 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml
```

Send a text request to the running server.

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"QwenPaw-Flash-9B","messages":[{"role":"user","content":"Hello, QwenPaw!"}],"max_tokens":128}'
```

## Family evaluation

The published QwenPaw evaluation covers five agent workflows. The figures report the original model-family comparisons.

![QwenPaw-Flash-9B comparison](https://img.alicdn.com/imgextra/i2/O1CN01uQsLX61XbwkjLtHuP_!!6000000002943-2-tps-1264-888.png)

![QwenPaw-Flash 2B, 4B and 9B compared with their base models](https://img.alicdn.com/imgextra/i1/O1CN01OKDCHQ1yuic00Qmyu_!!6000000006639-2-tps-1984-975.png)

## License

Apache License 2.0.