README.md

---
license: other
license_name: qwen-research
base_model: Qwen/Qwen2.5-3B-Instruct
library_name: rkllm
pipeline_tag: text-generation
tags:
- rkllm
- rockchip
- rk3576
- rk3588
- qwen
- qwen2.5

---

# Qwen2.5-3B-Instruct-RKLLM

RKLLM-converted versions of Qwen2.5-3B-Instruct for Rockchip RK3576 and RK3588.

`.rkllm` files require the Rockchip RKLLM runtime and cannot be loaded directly with Transformers, llama.cpp, or Ollama.

## Base Model

- Model: Qwen2.5-3B-Instruct
- Author: Alibaba Cloud Qwen Team
- Original Model: [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct)
- Original License: Qwen Research License

Refer to the [upstream model card](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) for the original model’s capabilities and limitations.

## Supported Platforms

| SoC    | Quantization | Status                   |
| ------ | ------------ | ------------------------ |
| RK3576 | W4A16        | ✅ Tested                 |
| RK3576 | W8A8         | ✅ Tested                 |
| RK3588 | W8A8         | ✅ Supported (not tested) |

RK3576 W4A16 and W8A8 have been tested. RK3588 W8A8 is supported as a target platform but has not been board-tested because no RK3588 board is available.

## Conversion

- RKLLM Toolkit: 1.2.3
- Target Platforms: RK3576, RK3588

## Available Variants

| File                                                         | Quantization |
| ------------------------------------------------------------ | ------------ |
| [Qwen2.5-3B-Instruct_w4a16_RK3576.rkllm](RK3576/Qwen2.5-3B-Instruct_w4a16_RK3576.rkllm) | W4A16        |
| [Qwen2.5-3B-Instruct_w8a8_RK3576.rkllm](RK3576/Qwen2.5-3B-Instruct_w8a8_RK3576.rkllm) | W8A8         |
| [Qwen2.5-3B-Instruct_w8a8_RK3588.rkllm](RK3588/Qwen2.5-3B-Instruct_w8a8_RK3588.rkllm) | W8A8         |

## Repository Structure

```text
.
├── Qwen2.5-3B-Instruct_data_quant.json
├── RK3576
│   ├── Qwen2.5-3B-Instruct_w4a16_RK3576.rkllm
│   └── Qwen2.5-3B-Instruct_w8a8_RK3576.rkllm
└── RK3588
    └── Qwen2.5-3B-Instruct_w8a8_RK3588.rkllm
```

## Usage

Download the desired RKLLM model and deploy it with the RKLLM runtime on the matching Rockchip platform.

For faster deployment without setting up the environment manually, see [Hanzo-Huang/rkllm-docker](https://github.com/Hanzo-Huang/rkllm-docker).

### Docker Deployment

Use the model-specific image for the fastest RK3576 W4A16 deployment. The model is included in the image:

```bash
sudo docker run --rm -it \
  --privileged \
  -p 8001:8001 \
  -v /dev:/dev \
  ghcr.io/hanzo-huang/rkllm-docker/qwen2.5-3b-instruct:w4a16-rk3576
```

Alternatively, use the reusable environment image and mount a downloaded model:

```bash
sudo docker run --rm -it \
  --privileged \
  -p 8001:8001 \
  -v /dev:/dev \
  -v ./:/app/models:ro \
  -e MODEL_PATH=/app/models/RK3576/Qwen2.5-3B-Instruct_w4a16_RK3576.rkllm \
  -e TARGET_PLATFORM=rk3576 \
  ghcr.io/hanzo-huang/rkllm-docker:env-latest
```

The second command requires downloading the selected `.rkllm` file first. Use the matching model variant and target platform.

## License

The original Qwen2.5-3B-Instruct model is released under the Qwen Research License. Converted artifacts are derived from the original model and remain subject to the applicable upstream license terms.

## Acknowledgements

Thanks to the Alibaba Cloud Qwen Team for releasing the original model and to Rockchip and RKLLM contributors for the deployment toolchain.