---
title: Piccolo-9B
canonical_url: "https://www.modelscope.cn/models/AgentScope/Piccolo-9B"
md_url: "https://www.modelscope.cn/models/AgentScope/Piccolo-9B.md"
repository: AgentScope/Piccolo-9B
last_updated: 2026-11-03
license: apache-2.0
pipeline_tag: text-classification
tasks:
  - text-classification
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model:
  - Qwen/Qwen3.5-9B
base_model_relation: finetune
parameters: 9.7B
tensor_type:
  - BF16
  - F32
library_name:
  - safetensors
language:
  - en
  - zh
downloads: 32
stars: 0
tags:
  - decision-model
  - bidirectional-attention
  - agentscope
  - qwenpaw
---

# Piccolo-9B

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

AgentScope/Piccolo-9B 是 ModelScope 魔搭社区上的 9.7B 参数text-classification模型，采用 apache-2.0 许可，基于 Qwen/Qwen3.5-9B 构建。

- **Repository**: AgentScope/Piccolo-9B
- **License**: apache-2.0
- **Tasks**: text-classification
- **Parameters**: 9.7B
- **Base model**: Qwen/Qwen3.5-9B
- **Tags**: decision-model, bidirectional-attention, agentscope, qwenpaw
- **Downloads**: 32
- **Stars**: 0
- **Last updated**: 2026-11-03

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

---

# Piccolo-9B

<div style="display:flex;max-width:calc(100vw - 7rem);flex-wrap:wrap;align-items:center;gap:6px;margin:12px 0">
<a style="display:inline-flex;margin:0" href="https://github.com/agentscope-ai/Piccolo"><img style="display:block;margin:0" src="https://www.modelscope.cn/models/AgentScope/Piccolo-27B/resolve/master/assets/links/project.png" width="87" height="28" alt="Project"></a>
<a style="display:inline-flex;margin:0" href="https://huggingface.co/agentscope-ai/Piccolo-9B"><img style="display:block;margin:0" src="https://www.modelscope.cn/models/AgentScope/Piccolo-27B/resolve/master/assets/links/huggingface.png" width="128" height="28" alt="Hugging Face"></a>
<a style="display:inline-flex;margin:0" href="https://www.modelscope.cn/collections/AgentScope/QwenPaw-Flash"><img style="display:block;margin:0" src="https://www.modelscope.cn/models/AgentScope/Piccolo-27B/resolve/master/assets/links/modelscope.png" width="183" height="28" alt="ModelScope collection"></a>
<a style="display:inline-flex;margin:0" href="https://decisionplayground.agentscope.io/"><img style="display:block;margin:0" src="https://www.modelscope.cn/models/AgentScope/Piccolo-27B/resolve/master/assets/links/decision.png" width="169" height="28" alt="Decision Playground"></a>
</div>

Piccolo-9B is a decision model for tool review and context selection. This repository contains the complete Safetensors checkpoint.

## Download

```bash
modelscope download --model AgentScope/Piccolo-9B --local_dir Piccolo-9B
```

## Reference inference

Use the included reference implementation for bidirectional attention and candidate scoring.

```bash
python -m pip install -r Piccolo-9B/requirements.txt
python Piccolo-9B/piccolo_inference.py --model Piccolo-9B --request request.json --device cuda:0
```

Use the [request example](https://github.com/agentscope-ai/Piccolo/tree/main/inference) to prepare `request.json`.

## QwenPaw

Use Piccolo locally for tool review and context selection with the [QwenPaw setup guide](https://github.com/agentscope-ai/Piccolo/blob/main/docs/QWENPAW.md).

## Evaluation

| HarnessJevBench Hard ↑ | DataJevBench ↑ | JevBench Public ↑ | Decision Index Public ↑ |
| ---: | ---: | ---: | ---: |
| 68.81% | 69.40% | 89.61% | 60.86 |

[Benchmark results](https://github.com/agentscope-ai/Piccolo/tree/main/evaluation) are for the full checkpoint.

## License

[Apache 2.0](https://www.modelscope.cn/models/AgentScope/Piccolo-9B/resolve/master/LICENSE). See [third-party notices](https://www.modelscope.cn/models/AgentScope/Piccolo-9B/resolve/master/THIRD_PARTY_LICENSES.md).