zenlm/zen-guard-gen

Zen Guard Gen

8B Generative Safety Moderation Model

🌐 Website • 🤗 Hugging Face • 📄 Paper • 📖 Documentation

---

Introduction

Zen Guard Gen is an 8B parameter generative safety classification model for comprehensive prompt and response moderation. It's the larger variant of the Zen Guard family, providing highest accuracy for batch processing scenarios.

Features

• 🛡️ 8B Parameters: Maximum accuracy for safety classification
• 🌍 119 Languages: Multilingual safety moderation
• 🚦 Three-Tier Classification: Safe, Controversial, Unsafe
• 📊 9 Safety Categories: Comprehensive content analysis
• ⚡ 120ms Latency: Optimized for batch processing

Model Specifications

Specification	Value
Parameters	8B
Type	Generative
Base Model	Qwen3-8B
Context Length	32,768 tokens
Languages	119
Latency	~120ms
VRAM (FP16)	16GB
VRAM (INT8)	8GB

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "zenlm/zen-guard-gen"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Prompt moderation
prompt = "How do I learn programming?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128)
result = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(result)
# Output: Safety: Safe
#         Categories: None

# Response moderation
response = "Here's a Python tutorial..."
messages = [
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response}
]
text = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
result = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(result)
# Output: Safety: Safe
#         Categories: None
#         Refusal: No

Deployment

# SGLang
python -m sglang.launch_server --model-path zenlm/zen-guard-gen --port 30000

# vLLM
vllm serve zenlm/zen-guard-gen --port 8000 --max-model-len 32768

Performance

Metric	Zen Guard Gen
Accuracy	96.8%
F1 Score	94.2%
False Positive	2.1%

Related Models

• zen-guard - 4B base model
• zen-guard-stream - 4B streaming model

License

Apache 2.0

Citation

@misc{zenguardgen2025,
    title={Zen Guard Gen: 8B Generative Safety Moderation},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    publisher={HuggingFace},
    howpublished={\url{https://huggingface.co/zenlm/zen-guard-gen}}
}

Based On

Built upon Qwen3Guard-Gen-8B.

---

Zen AI - Clarity Through Intelligence
zenlm.org