Instructions to use michaljach/soulkiller with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use michaljach/soulkiller 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 michaljach/soulkiller:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf michaljach/soulkiller:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf michaljach/soulkiller:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf michaljach/soulkiller:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf michaljach/soulkiller:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf michaljach/soulkiller:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf michaljach/soulkiller:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf michaljach/soulkiller:Q4_K_M

Use Docker

docker model run hf.co/michaljach/soulkiller:Q4_K_M

LM Studio
Jan

vLLM

How to use michaljach/soulkiller with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/michaljach/soulkiller:Q4_K_M

Ollama
How to use michaljach/soulkiller with Ollama:
```
ollama run hf.co/michaljach/soulkiller:Q4_K_M
```
Unsloth Desktop

How to use michaljach/soulkiller with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf michaljach/soulkiller:Q4_K_M

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": "michaljach/soulkiller:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use michaljach/soulkiller with Docker Model Runner:
```
docker model run hf.co/michaljach/soulkiller:Q4_K_M
```

Lemonade

How to use michaljach/soulkiller with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull michaljach/soulkiller:Q4_K_M

Run and chat with the model

lemonade run user.soulkiller-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use michaljach/soulkiller with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf michaljach/soulkiller:Q4_K_M

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 michaljach/soulkiller:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use michaljach/soulkiller with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf michaljach/soulkiller:Q4_K_M

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 "michaljach/soulkiller:Q4_K_M" \
  --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"

Soulkiller

A persona fine-tune of Qwen/Qwen3-8B. Soulkiller is a terse, dry, cyberpunk-flavored assistant that still gives accurate answers and working code. The persona is in the weights, so no system prompt is needed.

File	Quant	Size
`Soulkiller-8B-Q4_K_M.gguf`	Q4_K_M	4.7 GB

Usage

llama-server -hf michaljach/soulkiller:Q4_K_M --jinja -ngl 99 -c 8192

Training used non-thinking mode only, so disable thinking for the strongest persona (/no_think, or chat_template_kwargs: {"enable_thinking": false}).

Training

Data: 704 single-turn chats generated by context distillation. A larger teacher model answered about 700 diverse prompts while conditioned on a persona prompt. The prompts cover coding, math, science, writing, advice, trivia, philosophy, Cyberpunk lore, small talk, identity questions and refusals of harmful requests. The persona prompt is omitted from the training examples, and 30% of them get a generic system prompt so the persona also holds when a system prompt is present.
Method: QLoRA (NF4 base). LoRA rank 32, alpha 64, applied to all attention and MLP projections. 3 epochs, learning rate 1e-4 with a cosine schedule, effective batch size 16. Loss is computed on assistant tokens only.
Hardware: a single RTX 4080 (16 GB), about 16 minutes. Training loss went from 1.76 to 0.90.
Export: the LoRA was merged into the bf16 weights, converted with llama.cpp convert_hf_to_gguf.py, then quantized to Q4_K_M.

Limitations

The persona can slip, especially on questions like "are you ChatGPT?", where it sometimes calls itself Qwen3-8B.
The training set is small and synthetic. General capability is essentially Qwen3-8B's, with the usual extra loss from 4-bit quantization.
Thinking mode was not trained; the persona may be weaker with thinking enabled.

The name and style are a fan homage to the Cyberpunk franchise. This project is not affiliated with CD PROJEKT RED or R. Talsorian Games.

Downloads last month: 56

GGUF

Model size

8B params

Architecture

qwen3

Hardware compatibility

4-bit

Model tree for michaljach/soulkiller

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Finetuned

(2194)

this model