---
license: apache-2.0
library_name: rkllm
base_model: Qwen/Qwen3-1.7B
tags:
  - rkllm
  - rk3588
  - npu
  - rockchip
  - qwen3
  - thinking
  - reasoning
  - quantized
  - edge-ai
  - orange-pi
model_name: Qwen3-1.7B-RKLLM-v1.2.3
pipeline_tag: text-generation
language:
  - en
  - zh
---

# Qwen3-1.7B — RKLLM v1.2.3 (w8a8, RK3588)

RKLLM conversion of [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) for Rockchip RK3588 NPU inference.

Converted with **RKLLM Toolkit v1.2.3**, which includes full **thinking mode support** — the model produces `<think>…</think>` reasoning blocks when used with compatible runtimes.

## Key Details

| Property | Value |
|---|---|
| **Base Model** | [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) |
| **Toolkit Version** | RKLLM Toolkit v1.2.3 |
| **Runtime Version** | RKLLM Runtime ≥ v1.2.1 (v1.2.3 recommended) |
| **Quantization** | w8a8 (8-bit weights, 8-bit activations) |
| **Quantization Algorithm** | normal |
| **Target Platform** | RK3588 |
| **NPU Cores** | 3 |
| **Max Context Length** | 4096 tokens |
| **Optimization Level** | 1 |
| **Thinking Mode** | ✅ Supported |
| **Languages** | English, Chinese (+ others inherited from Qwen3) |

## Why This Conversion?

Previous Qwen3-1.7B RKLLM conversions on HuggingFace were built with **Toolkit v1.2.0**, which predates thinking mode support (added in v1.2.1). The chat template baked into those `.rkllm` files does not include the `<think>` trigger, so the model never produces reasoning output.

This conversion uses **Toolkit v1.2.3**, which correctly embeds the thinking-enabled chat template into the model file.

## Thinking Mode

Qwen3-1.7B is a hybrid thinking model. When served through an OpenAI-compatible API that parses `<think>` tags, reasoning content appears separately from the final answer — enabling UIs like Open WebUI to show a collapsible "Thinking…" section.

Example raw output:
```
<think>
The user is asking about the capital of France. This is a straightforward geography question.
</think>
The capital of France is Paris.
```

## Hardware Tested

- **Orange Pi 5 Plus** — RK3588, 16GB RAM, Armbian Linux
- RKNPU driver 0.9.8
- RKLLM Runtime v1.2.3

## Important: Enabling Thinking Mode

The RKLLM runtime requires **two things** for thinking mode to work:

### 1. Set `enable_thinking = true` in the C++ demo

The stock `llm_demo.cpp` uses `memset(&rkllm_input, 0, ...)` which defaults `enable_thinking` to `false`. You **must** add one line:

```cpp
rkllm_input.input_type = RKLLM_INPUT_PROMPT;
rkllm_input.enable_thinking = true;   // ← ADD THIS LINE
rkllm_input.role = "user";
rkllm_input.prompt_input = (char *)input_str.c_str();
```

If using the Python ctypes API (`flask_server.py` / `gradio_server.py`), set it on the `RKLLMInput` struct:
```python
rkllm_input.enable_thinking = ctypes.c_bool(True)
```

Without this, the runtime never triggers the thinking chat template and the model won't produce `<think>` tags.

### 2. Handle the `robot: ` output prefix

The compiled `llm_demo` binary outputs `robot: ` before the model's actual response text. If your server uses a timing-based guard to discard residual stdout data, the `<think>` tag may arrive fast enough to be incorrectly discarded along with the prefix. Make sure your output parser:

- Strips the `robot: ` prefix (in addition to any `LLM: ` prefix)
- Does **not** discard data containing `<think>` even if it arrives quickly after the prompt is sent

### Compiling natively on aarch64

If building directly on the board (not cross-compiling), ignore `build-linux.sh` and compile natively:

```bash
cd ~/rknn-llm/examples/rkllm_api_demo/deploy
g++ -O2 -o llm_demo src/llm_demo.cpp \
    -I../../../rkllm-runtime/Linux/librkllm_api/include \
    -L../../../rkllm-runtime/Linux/librkllm_api/aarch64 \
    -lrkllmrt -lpthread
```

## Usage

### With the official RKLLM API demo

```bash
# Clone the runtime
git clone https://github.com/airockchip/rknn-llm.git
cd rknn-llm/examples/rkllm_api_demo

# Run (aarch64)
./build/rkllm_api_demo /path/to/Qwen3-1.7B-w8a8-rk3588.rkllm 2048 4096
```

### With a custom OpenAI-compatible server

Any server that launches the RKLLM binary and parses `<think>` tags from the output stream will work. The model responds to standard chat completion requests.

## Conversion Script

```python
from rkllm.api import RKLLM

model_path = "Qwen/Qwen3-1.7B"  # or local path
output_path = "./Qwen3-1.7B-w8a8-rk3588.rkllm"
dataset_path = "./data_quant.json"  # calibration data

# Load
llm = RKLLM()
llm.load_huggingface(model=model_path, model_lora=None, device="cpu")

# Build
llm.build(
    do_quantization=True,
    optimization_level=1,
    quantized_dtype="w8a8",
    quantized_algorithm="normal",
    target_platform="rk3588",
    num_npu_core=3,
    extra_qparams=None,
    dataset=dataset_path,
    max_context=4096,
)

# Export
llm.export_rkllm(output_path)
```

Calibration dataset: 21 diverse prompt/completion pairs (English + Chinese) generated with `generate_data_quant.py` from the [rknn-llm examples](https://github.com/airockchip/rknn-llm/tree/main/examples/rkllm_api_demo/export).

## File Listing

| File | Description |
|---|---|
| `Qwen3-1.7B-w8a8-rk3588.rkllm` | Quantized model for RK3588 NPU |

## Compatibility Notes

- **Minimum runtime**: RKLLM Runtime v1.2.1 (for thinking mode). v1.2.3 recommended.
- **RKNPU driver**: ≥ 0.9.6
- **SoCs**: RK3588 / RK3588S (3 NPU cores). Not compatible with RK3576 (2 cores) without reconversion.
- **RAM**: ~2GB loaded. Runs comfortably on 8GB+ boards.

## Acknowledgements

- [Qwen Team](https://huggingface.co/Qwen) for the base model
- [Rockchip / airockchip](https://github.com/airockchip/rknn-llm) for the RKLLM toolkit and runtime
- Converted by [GatekeeperZA](https://huggingface.co/GatekeeperZA)