Qwen3Guard - Ali Qwen open source security model

What is Qwen3Guard

Qwen3Guard is based on Qwen3 Base model fine-tuned security protection model designed for security detection. Accurate security categorization of prompts and responses, risk levels, and support for English, Chinese, and multi-language environments.Qwen3Guard is available in two specialized variants: Qwen3Guard-Gen for offline security annotation and filtering of datasets, and Qwen3Guard-Stream for real-time streaming of security inspections, which enables instant review of content as the model generates responses. The former is suitable for offline security labeling and filtering datasets, and the latter enables real-time streaming security inspection, which can instantly review content during the model-generated response process. Both models are available in 0.6B, 4B and 8B sizes to accommodate different deployment scenarios and resource constraints, and Qwen3Guard's core highlights include real-time streaming detection technology, three levels of risk classification (safe, unsafe, controversial), and multi-language support (covering 119 languages). It performs well in major security benchmarks and is suitable for a wide range of deployment scenarios.

Features of Qwen3Guard

• Precise safety classification: Can perform precise security testing of cues and responses, providing risk levels and categorization to ensure content security.
• Real-time streaming detection: Qwen3Guard-Stream supports real-time security detection during model response generation, ensuring low latency and high efficiency.
• Multi-language support: Support for 119 languages and dialects for global and cross-language scenarios.
• Tertiary risk levelProvides three labels of "safe", "unsafe" and "controversial", allowing users to flexibly adjust their security policies according to their needs.
• Open Source and Ease of Use: Models can be downloaded from Hugging Face or ModelScope, and are supported to be used through the AliCloud AI guardrail service for easy deployment and application.

Core Benefits of Qwen3Guard

• Efficient real-time detectionQwen3Guard-Stream can perform real-time security detection during the process of generating responses, ensuring content security without sacrificing response speed, suitable for scenarios with high real-time requirements.
• Flexible risk categorizationIt provides three levels of risk categorization, namely "secure", "insecure" and "controversial", so that users can flexibly adjust the security policy according to their specific needs and adapt to different application scenarios.
• Powerful multi-language support: Supporting 119 languages and dialects, it has wide applicability globally and can meet the security detection needs in cross-lingual environments.
• Advanced technical performance: It performs well in major security benchmark tests, demonstrating its strong capabilities in security classification tasks and providing users with reliable protection.

What is Qwen3Guard's official website?

• Project website:: https://qwen.ai/blog?id=f0bbad0677edf58ba93d80a1e12ce458f7a80548&from=research.research-list
• Github repository:: https://github.com/QwenLM/Qwen3Guard
• HuggingFace Model Library:: https://huggingface.co/collections/Qwen/qwen3guard-68d2729abbfae4716f3343a1
• Technical Report:: https://github.com/QwenLM/Qwen3Guard/blob/main/Qwen3Guard_Technical_Report.pdf

People for whom Qwen3Guard is intended

• Enterprise Security Team: The output of generative AI needs to be monitored and audited in real time to ensure that the content meets enterprise security standards and compliance requirements.
• Content Review Organization: Responsible for performing security audits of large amounts of textual content, efficient and accurate tools are needed to supplement the manual review process.
• AI developers and researchers: During the development and research process, the generated text content needs to be evaluated for security to optimize the performance and security of the model.
• Social media platforms: User-generated content needs to be monitored in real time to prevent the spread of harmful information and maintain a healthy environment on the platform.
• educational organization: When using AI-assisted instruction, there is a need to ensure that the content generated is appropriate for the student population and to avoid inappropriate content.
• Government and regulatory bodies: AI-generated content needs to be regulated to ensure that it complies with laws, regulations and social and ethical standards.