Loading page…
Qwen3Guard offers complete-message and streaming moderation, with safe, controversial and unsafe labels.
Qwen3Guard makes the timing of moderation an explicit choice. A product can inspect complete content or evaluate an evolving response. That choice affects delay, what a reviewer sees and whether anything has already reached the user when a decision changes.
The Qwen team describes Gen for classification of complete prompts and responses, and Stream for incremental token-level checks. Both are offered in 0.6B, 4B and 8B sizes.
Source: Qwen team · Qwen3Guard: Gen and StreamThe documentation distinguishes safe, controversial and unsafe outputs and describes multilingual coverage. A product still needs to decide how each label maps to its policy and to evaluate its own language mix.
Source: Qwen team · Qwen3Guard: Gen and StreamOur recommendation: explicitly choose whether to buffer output, display checked chunks or allow immediate streaming. Stopping a stream cannot retract text already read by a user or consumed by another system.
For a live writing assistant, compare complete-answer review with delayed chunk display using the same examples. Record the time to useful text, unsafe text exposed before interruption, and harmless text incorrectly stopped.
Streaming availability does not mean zero delay or zero exposure. Published language breadth is not evidence of uniform performance, and the controversial label is not a universal social judgment.
How to read an AI safety evaluationThe sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
Explains complete-message and incremental moderation, model sizes and the three-level safety labels.
A framework for identifying, measuring and managing generative AI risks across the system lifecycle.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence