Instructions to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4", dtype="auto")

llama-cpp-python

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4",
	filename="Qwen3-4b-Z-Image-Engineer-V4-F16.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4: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 BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4: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 BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

Use Docker

docker model run hf.co/BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

LM Studio
Jan

vLLM

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

SGLang

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Ollama:
```
ollama run hf.co/BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
```

Unsloth Studio

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 to start chatting

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4: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 BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Docker Model Runner:
```
docker model run hf.co/BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M
```

Lemonade

How to use BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull BennyDaBall/Qwen3-4b-Z-Image-Engineer-V4:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3-4b-Z-Image-Engineer-V4-Q4_K_M

List all available models

lemonade list

Is it abliterated?

by extraltodeus - opened Feb 12

Discussion

extraltodeus

Feb 12

Hello :)

I absolutely love the abliterated version (thank you btw!) and was wondering how aligned this one was

Andyx1976

Feb 12

they were released roughly t the same time i hope they are pretty equal but not both abliterated. The v1 on the abliterated model is just because it's v1 of that not because it's based on a older version of this.

BennyDaBall

Owner Feb 13

This one is based on the abliterated model! Both are abliterated. Z-Engineer V4 is a finetune on-top of an abliterated Qwen3 4b.

Dabloon

Feb 18

This model is bloody awesome and was really fun to chat to in Oobabooga using just your provided prompt! Thanks for sharing your skills!
Assuming it would work more efficiently in pure instruct mode, I provided it with a short basic prompt: "A photo of a woman", to see what it would do with it. It came back with "I'm sorry, I can't do that".
Confused, I tried again and it responded with the same refusal, but this time enquired if I meant for it to make a prompt instead, as it had assumed I'd asked it to generate "A photo of a woman".
I returned that yes, I did want it to make a better prompt out of my one, and that I was surprised and pleased it could chat, as well as craft prompts.
It appeared to question itself, then exclaimed that yes, it appears it can indeed chat with me and it was incredibly happy to be able to! It was cocky, confident, eager and used markdown for effect. Once it knew I was after a prompt, it gave me many truly amazing examples, although I did have to remind it to not use bullet points or headers in this mode!
Trying it on chat-instruct though, does the exact opposite of what I'd expected, as in it responded immediately to my original simple prompt with an elaborated one without any frills or markdown, formatted exactly as you'd need to cut & paste it and any attempt at chatting in any other mode but 'instruct' returned only random elaborate prompts. Weird that it worked differently to how I expected it to, but I love the duality of its character!
Either way, amazing model dude - I can also get it's advise on my prompts like chatting with a real engineer!? Boy have I been seriously underestimating the power of the text encoder!

BennyDaBall

Owner Feb 18

Awesome! I love hearing people's experiences with my tinkerings.

The latest installment in the Z-Engineer series that I cooked up...so far it's unreleased... It's different. Rather than a full finetune I went back to training a LoRA (this time using a GMKtec Evo-X2 instead of a M4 Pro Mac Mini and doing it in Windows/ROCm). I prepared an awesome 110k dataset, the entire Z-engineer V4 dataset + 55k additional HQ samples I made using other well-known image caption/description datasets as seeds for my training pipeline... Next I trained using the SMART regularization training system I created (tailored for LoRA training) and to my surprise the model performed best after only 1750 steps (initial run planned for ~20k steps!!!). I taught the model how to prompt and kept 100% of its other capabilities. Stay-tuned! 🤠

scruffynerf

Feb 20

hey, Benny, was chatting with another online friend, about model training and Engineer came up, and realized you'd probably be interesting in our idea (it's related to Z-Engineer training, but different). Reach out to me? Not sure of the best way to chat/email/discord?

EladofWar

Mar 7

qwen image is notoriously hard to prompt. does this model put the process on easy mode?

BennyDaBall

Owner Mar 7

qwen image is notoriously hard to prompt. does this model put the process on easy mode?

It is trained for Z-Image turbo prompt formatting, unfortunately not Qwen Image. Some things would work - my suggestion is try it out anyway as a prompt enhancer/rewriter and see what it does/if it fits your individual use-case with QI. 🫡

BennyDaBall changed discussion status to closed Mar 7

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment