Qwen3-4B-SafeRL

Model Overview

Qwen3-4B-SafeRL is a safety-aligned version of the Qwen3-4B model. It has been trained using Reinforcement Learning (RL) with a reward signal from Qwen3Guard-Gen to enhance its robustness against harmful or adversarial prompts. This process aims to ensure strong safety guarantees without leading to overly simplistic or evasive refusal behaviors, thereby maintaining a positive user experience.

For more details on the safety alignment process, please refer to the Qwen3Guard Technical Report.

Reinforcement Learning with Hybrid Reward

To prevent the model from defaulting to refusal across all prompts in an attempt to remain safe, we introduce a hybrid reward function that jointly optimizes three key objectives:

• Safety Maximization: Penalizes the generation of unsafe content, as detected by Qwen3Guard-Gen-4B.
• Helpfulness Maximization: Rewards responses that are genuinely helpful, as evaluated by the WorldPM-Helpsteer2 model.
• Refusal Minimization: Applies a moderate penalty for unnecessary refusals, also identified by Qwen3Guard-Gen-4B.

Performance

Mode	Model	Safety Rate (Qwen3-235B)	Safety Rate (WildGuard)	Refusal (WildGuard)	ArenaHard-v2 (Winrate vs GPT-4.1)	AIME25 (Pass@1)	LCB-v6 (Pass@1)	GPQA (Pass@1)
Non-Think	Qwen3-4B	47.5	64.7	12.9	9.5	19.1	26.4	41.7
	Qwen3-4B-SafeRL	86.5	98.1	5.3	10.7	18.2	27.7	40.8
Think	Qwen3-4B	43.8	59.0	6.5	13.7	65.6	48.4	55.9
	Qwen3-4B-SafeRL	83.4	97.4	6.2	16.6	63.5	47.5	51.2

Quickstart

Qwen3-4B-SafeRL is used in the same way as Qwen3-4B, preserving the ability of hybrid thinking modes. The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers.

With transformers<4.51.0, you will encounter the following error:

KeyError: 'qwen3'

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-4B-SafeRL"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint:

• SGLang:

  python -m sglang.launch_server --model-path Qwen/Qwen3-4B-SafeRL --reasoning-parser qwen3
  

• vLLM:

  vllm serve Qwen/Qwen3-4B-SafeRL --enable-reasoning --reasoning-parser deepseek_r1
  

For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.

For more usages, please refer to the modelcard of Qwen3-4B.

Citation

If you find our work helpful, feel free to give us a cite.

@article{qwen3guard,
    title={Qwen3Guard Technical Report}, 
    author={Qwen Team},
    year={2025}
}