jamescallander/DeepSeek-R1-Distill-Llama-8B_w8a8_g128_rk3588.rkllm

DeepSeek-R1-Distill-Llama-8B — RKLLM build for RK3588 boards

Built with DeepSeek (DeepSeek License Agreement)

Author: @jamescallander
Source model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B · Hugging Face
Target: Rockchip RK3588 NPU via RKNN-LLM Runtime

This repository hosts a conversion of DeepSeek-R1-Distill-Llama-8B for use on Rockchip RK3588 single-board computers (Orange Pi 5 plus, Radxa Rock 5b+, Banana Pi M7, etc.). Conversion was performed using the RKNN-LLM toolkit

Conversion details

• RKLLM-Toolkit version: v1.2.1
• NPU driver: v0.9.8
• Python: 3.12
• Quantization: w8a8_g128
• Output: single-file .rkllm artifact
• Tokenizer: not required at runtime (UI handles prompt I/O)

Intended use

• On-device inference on RK3588 SBCs.
• Reasoning-focused model — designed to handle multi-step thinking, problem-solving, and structured explanations.
• Well-suited for tasks that need step-by-step reasoning or more careful breakdowns than typical instruction models.

Limitations

• Requires 9GB free memory
• Quantized build (w8a8_g128) may show small quality differences vs. full-precision upstream.
• Tested on Radxa Rock 5B+; other devices may require different drivers/toolkit versions.
• While strong at reasoning, performance is limited by RK3588’s NPU compared to high-end GPUs.

Quick start (RK3588)

1) Install runtime

The RKNN-LLM toolkit and instructions can be found on the specific development board's manufacturer website or from airockchip's github page.

Download and install the required packages as per the toolkit's instructions.

2) Simple Flask server deployment

The simplest way the deploy the .rkllm converted model is using an example script provided in the toolkit in this directory: rknn-llm/examples/rkllm_server_demo

python3 <TOOLKIT_PATH>/rknn-llm/examples/rkllm_server_demo/flask_server.py \
  --rkllm_model_path <MODEL_PATH>/DeepSeek-R1-Distill-Llama-8B_w8a8_g128_rk3588.rkllm \
  --target_platform rk3588

3) Sending a request

A basic format for message request is:

{
    "model":"DeepSeek-R1-Distill-Llama-8B",
    "messages":[{
        "role":"user",
        "content":"<YOUR_PROMPT_HERE>"}],
    "stream":false
}

Example request using curl:

curl -s -X POST <SERVER_IP_ADDRESS>:8080/rkllm_chat \
    -H 'Content-Type: application/json' \
    -d '{"model":"DeepSeek-R1-Distill-Llama-8B","messages":[{"role":"user","content":"In 2 or 3 sentences, who was Napoleon Bonaparte?"}],"stream":false}'

The response is formated in the following way:

{
    "choices":[{
        "finish_reason":"stop",
        "index":0,
        "logprobs":null,
        "message":{
            "content":"<MODEL_REPLY_HERE">,
            "role":"assistant"}}],
        "created":null,
        "id":"rkllm_chat",
        "object":"rkllm_chat",
        "usage":{
            "completion_tokens":null,
            "prompt_tokens":null,
            "total_tokens":null}
}

Example response:

{"choices":[{"finish_reason":"stop","index":0,"logprobs":null,"message":{"content":"","role":"assistant"}}],"created":null,"id":"rkllm_chat","object":"rkllm_chat","usage":{"completion_tokens":null,"prompt_tokens":null,"total_tokens":null}} 

Note on reasoning traces

This model outputs intermediate reasoning text (e.g., chains of thought) before its final response, enclosed by </think> markers.

• Many OpenAI-compatible UIs automatically suppress or hide this internal reasoning.
• If your client does not, you may see the reasoning steps along with the final answer.

4) UI compatibility

This server exposes an OpenAI-compatible Chat Completions API.

You can connect it to any OpenAI-compatible client or UI (for example: Open WebUI)

• Configure your client with the API base: http://<SERVER_IP_ADDRESS>:8080 and use the endpoint: /rkllm_chat
• Make sure the model field matches the converted model’s name, for example:

{
 "model": "DeepSeek-R1-Distill-Llama-8B",
 "messages": [{"role":"user","content":"Hello!"}],
 "stream": false
}

License

This conversion follows the MIT License

• Attribution: Built with DeepSeek-R1-Distill-Llama-8B (DeepSeek-AI)
• Required notice: see NOTICE
• Modifications: quantization (w8a8_g128), export to .rkllm format for RK3588 SBCs