Tongyi Qianwen has launched the Qwen3Guard security review model series, featuring cross-language, real-time, and implementable features. Supporting 119 languages and dialects, the series offers three parameter scales (0.6B, 4B, and 8B) and two form factors: Qwen3Guard-Stream for low-latency streaming detection (for real-time conversations, live broadcasts, and online applications); and Qwen3Guard-Gen for full-context security analysis and policy interpretation, suitable for offline assessment scenarios such as reinforcement learning reward modeling. Model output utilizes a three-level risk classification system: Safe/Controversial/Unsafe , along with multi-dimensional category labels to facilitate policy implementation and auditing.
Official technical materials and community pages indicate that this series achieves leading or comparable performance on English, Chinese, and multilingual security benchmarks. Hugging Face and ModelScope weighting and inference examples are available, along with a technical report and usage guide. Actual effectiveness depends on deployment latency, thresholds, and scenario customization. For highly sensitive or compliance-critical scenarios, manual review and business blacklisting are still recommended.
Frequently Asked Questions
Q: What models and uses are there?
A: There are three speeds: 0.6B/4B/8B. Stream is suitable for real-time low-latency review, and Gen is suitable for full context judgment and RL reward modeling.
Q: What languages are supported?
A: Covering 119 languages and dialects, emphasizing cross-language robustness and handling of ambiguous text and spoken language variants.
Q: How to interpret the output?
A: It provides risk classification (Safe/Controversial/Unsafe) and category labels, which can be used to map to interception, demotion, or manual review.
Q: Is it open source?
A: We provide open-source weights and inference examples, including configuration and inference scripts; the technical report details the data and training details.
Q: How to integrate into existing systems?
A: Load weights according to warehouse examples, set thresholds and category mappings; use Qwen3Guard-Stream for streaming scenarios and Gen for offline/training scenarios.
