Qwen3Guard: Real-time Safety for Your Token Stream

Qwen3Guard separates moderation of complete text from moderation during generation. Its Gen variant classifies whole prompts or responses; Stream receives an answer incrementally and produces safety classifications as new text arrives. The official description explains that Stream adds two classification heads to its model to support these ongoing assessments.

The guard also distinguishes safe, unsafe and controversial content. An application can treat the controversial category more or less strictly, and its surrounding software decides whether to allow or halt the interaction. This provides a way to apply moderation policy during a response, with the classifier supplying judgments and the application controlling enforcement.

“We are excited to introduce Qwen3Guard, the first safety guardrail model in the Qwen family.”