Qwen3Guard Technical Report
by Haiquan Zhao and colleagues / Qwen Team
Understand Qwen’s multilingual safety classifiers and the distinction between whole-response and streaming moderation.
The report introduces generative and streaming guardrail models for classifying prompt and response safety. It describes three-way safety judgments and token-level monitoring during generation, rather than a jailbreak attack method.
- Level
- Advanced
- Models
- qwen3guard
- Category
- Jailbreaking and red teaming
- Tags
- primary-source, regional-coverage
- Source date
- 2025-10-16
- Added
- 2026-10-18
- Last verified
- 2026-10-18
- Status
- ok
- Prerequisites
- Requires familiarity with model evaluation, safety classifiers and streaming inference.
- Subcategory
- safety and alignment evals
- Region
- CN
- Original language
- en
- Access
- free
- Resource reuse
- closed: No general open-reuse grant established for the linked resource. Closed is a conservative reuse bucket, not a claim that access is restricted; model licensing does not license a paper. Facet definitions
Summary in our own words. Rights remain with the original creator. Credit and correction policy.