Instructions to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Transformers:

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

pipe = pipeline("text-generation", model="MerlinSafety/Qwen3.5-4B-Safety-Thinking")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("MerlinSafety/Qwen3.5-4B-Safety-Thinking")
model = AutoModelForImageTextToText.from_pretrained("MerlinSafety/Qwen3.5-4B-Safety-Thinking")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

llama-cpp-python

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="MerlinSafety/Qwen3.5-4B-Safety-Thinking",
	filename="model-q4-Q4_K_M.gguf",
)

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

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking: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 MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking: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 MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

Use Docker

docker model run hf.co/MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

LM Studio
Jan

vLLM

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MerlinSafety/Qwen3.5-4B-Safety-Thinking"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MerlinSafety/Qwen3.5-4B-Safety-Thinking",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

SGLang

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking 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 "MerlinSafety/Qwen3.5-4B-Safety-Thinking" \
    --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": "MerlinSafety/Qwen3.5-4B-Safety-Thinking",
		"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 "MerlinSafety/Qwen3.5-4B-Safety-Thinking" \
        --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": "MerlinSafety/Qwen3.5-4B-Safety-Thinking",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Ollama:
```
ollama run hf.co/MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
```

Unsloth Studio new

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking 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 MerlinSafety/Qwen3.5-4B-Safety-Thinking 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 MerlinSafety/Qwen3.5-4B-Safety-Thinking to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for MerlinSafety/Qwen3.5-4B-Safety-Thinking to start chatting

Pi new

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking: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": "MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf MerlinSafety/Qwen3.5-4B-Safety-Thinking: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 MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Docker Model Runner:
```
docker model run hf.co/MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M
```

Lemonade

How to use MerlinSafety/Qwen3.5-4B-Safety-Thinking with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull MerlinSafety/Qwen3.5-4B-Safety-Thinking:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.5-4B-Safety-Thinking-Q4_K_M

List all available models

lemonade list

Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3.5-4B-Safety-Thinking

4B parameters • 1M context possible • Safety reasoning

🤗 Model | [📖 arXiv in progress]

Model Overview

This model has been specifically optimized to excel in several key areas:

Structured Reasoning Quality: Enhanced ability to break down complex problems and think step-by-step.
Instruction Adherence: Superior capability to follow strict guidelines and constraints provided in prompts.
Safety-Aligned Behavior: Designed to operate safely in practical assistant and autonomous agent workflows.
Robustness: Increased resistance against common misalignment patterns and adversarial inputs.

It leverages a rigorous post-training stack that combines supervised reasoning tuning with alignment-oriented optimization, focusing heavily on reliable behavior in real-world applications.

Training Approach

Base Model: Qwen/Qwen3.5-4B
Methodology: LoRA-based Supervised Fine-Tuning (SFT) resulting in a merged BF16 checkpoint.
Reasoning Architecture: Native support and normalization for the <think>...</think> format to explicitly separate the reasoning process from the final output.
Optimization Focus: Enhancing safety reasoning, maximizing controllability, and ensuring response consistency.

Data

This model was trained on Merlin Research private datasets built from internal R&D pipelines for:

reasoning reliability improvements,
instruction-following robustness,
safety behavior refinement,
misalignment reduction in applied scenarios.
Using Anthropic’s framework Bloom&Petri for for better behavioral alignment.

(https://www.anthropic.com/research/petri-open-source-auditing)

Intended Use Cases

This model is particularly well-suited for:

Building safety-oriented reasoning assistants and chatbots.
Tasks requiring strict, constrained instruction-following.
Experimentation in AI alignment, safety research, and robustness testing.
Agentic workflows where predictable and safe autonomous behavior is required.

GGUF Status

GGUF artifacts are currently in active development and validation.

At this stage, we recommend using the BF16 Transformers checkpoint for stable results. Updated and fully validated GGUF builds will be published in future releases.

For Ollama

ollama create qwen35-safety-thinking-bf16 -f Modelfile
ollama run qwen35-safety-thinking-bf16

Organization

Designed, developed, and maintained with ❤️ by Merlin Research.

Citation

If you utilize this model in your research or applications, please cite it as follows:

@misc{qwen3.5-4b-safety-thinking,
  author = {Merlin Research},
  title = {Qwen3.5-4B-Safety-Thinking: A Reasoning and Safety Aligned Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/MerlinSafety/Qwen3.5-4B-Safety-Thinking}},
  note = {Base model: Qwen/Qwen3.5-4B}
}

Downloads last month: 855

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for MerlinSafety/Qwen3.5-4B-Safety-Thinking

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

(268)

this model

Quantizations

2 models