Instructions to use Harsha001/Raon-1-nano-ROCmFPX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Harsha001/Raon-1-nano-ROCmFPX with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0
# Run inference directly in the terminal:
llama cli -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0
# Run inference directly in the terminal:
llama cli -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Use Docker

docker model run hf.co/Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Llama new
LM Studio
Jan

vLLM

How to use Harsha001/Raon-1-nano-ROCmFPX with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Harsha001/Raon-1-nano-ROCmFPX"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Harsha001/Raon-1-nano-ROCmFPX",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Ollama
How to use Harsha001/Raon-1-nano-ROCmFPX with Ollama:
```
ollama run hf.co/Harsha001/Raon-1-nano-ROCmFPX:Q8_0
```
Unsloth Desktop

How to use Harsha001/Raon-1-nano-ROCmFPX with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Harsha001/Raon-1-nano-ROCmFPX:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Harsha001/Raon-1-nano-ROCmFPX with Docker Model Runner:
```
docker model run hf.co/Harsha001/Raon-1-nano-ROCmFPX:Q8_0
```

Lemonade

How to use Harsha001/Raon-1-nano-ROCmFPX with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Run and chat with the model

lemonade run user.Raon-1-nano-ROCmFPX-Q8_0

List all available models

lemonade list

Hermes Agent

How to use Harsha001/Raon-1-nano-ROCmFPX with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Harsha001/Raon-1-nano-ROCmFPX with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Harsha001/Raon-1-nano-ROCmFPX:Q8_0

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Harsha001/Raon-1-nano-ROCmFPX:Q8_0" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Raon-1 nano, ROCmFPX GGUF

An 8-bit GGUF of Raon-1 nano by KRAFTON AI, a small Korean and English model for conversation, instruction following and tool calling (3.1B parameters, of which 1.5B are a per-layer embedding table).

This file needs the raon-1-nano branch of HarshaVardhan3002/ROCmFPX. 23 of its weights use the ROCmFPX 8-bit type, so stock llama.cpp and KRAFTON's own llama.cpp branch will not load it. If you do not use ROCmFPX, take KRAFTON's official GGUFs: their Q8_0 has the same quality.

Not tested on AMD hardware yet. Please report speed and quality in the Community tab.

Not an official release. All credit for the model goes to KRAFTON AI.

File

File	Size	KL to F16	Same top token as F16
`Raon-1-nano-Q8_0-ROCmFPX-mix.gguf`	3.45 GB	0.000341	98.90%

For comparison, measured the same way:

Not published	Size	KL to F16	Same top token
Stock Q8_0	3.45 GB	0.000344	98.95%
ROCmFPX 8-bit on every main weight	3.42 GB	0.000363	98.84%
Stock Q6_K	2.70 GB	0.0034	96.2%
ROCmFPX 6-bit	2.70 GB	0.0037	95.8%
Stock Q4_K_M	2.13 GB	0.0197	91.7%
ROCmFPX 4-bit (FP4)	2.14 GB	0.0221	91.2%

What this file is: a stock Q8_0 where each weight was compared with its ROCmFPX 8-bit version, and the ROCmFPX version was kept wherever its error was lower. That was 23 weights: the attention output in 16 layers and the feed-forward down projection in 7 layers. The result is the same quality as stock Q8_0 (the difference is within the measurement error of about ±0.000004) and 4 MB smaller. It is not a big improvement. On this model the stock 6-bit and 4-bit types beat the ROCmFPX ones on every weight, so only the 8-bit file is published.

How this was measured: llama.cpp llama-perplexity --kl-divergence against an F16 GGUF converted from KRAFTON's weights (perplexity 19.20; this file 19.22), 40 chunks of 512 tokens of English Wikipedia text (wikitext-2), after expanding each file to F16. KL is how far the file's next-token probabilities drift from F16, averaged over every token; lower is better. The text is English, so it says nothing direct about Korean or tool calling.

Short checks on CPU with greedy decoding: it answered 17 x 23 = 391, named Canberra as Australia's capital, wrote a correct iterative Fibonacci function in Python, and answered a Korean question about Korea's capital correctly in Korean ("한국의 수도는 서울입니다.").

The file contains the model's one-layer multi-token prediction (MTP) head, which was not tested.

Running it

Build the raon-1-nano branch of ROCmFPX, then:

llama-server -m Raon-1-nano-Q8_0-ROCmFPX-mix.gguf -ngl 99 -c 32768 --jinja

Credits and license

Model: KRAFTON/Raon-1-nano, Apache-2.0.

Apache-2.0. See LICENSE.

Downloads last month: 22

GGUF

Model size

3B params

Architecture

raon1nano

Hardware compatibility

8-bit

Model tree for Harsha001/Raon-1-nano-ROCmFPX

Base model

KRAFTON/Raon-1-nano

Quantized

(5)

this model