Qwen3Guard (Gen and Stream; 0.6B, 4B, 8B)
Qwen's first safety guardrail models, Gen for prompt/response classification and Stream for token-level real-time moderation, at 0.6B, 4B and 8B.
Releases real-time guardrails that screen a model's output as it streams (Stream variants) as well as full prompts and responses (Gen variants). Open weights on Hugging Face.
- Date
- Tuesday, 23 September 2025
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Technical report | arXiv 2510.14276 Submitted 2025-10-16; tri-class (safe, controversial, unsafe) judgments and token-level streaming head | company |
Blog describes it as the first safety guardrail model in the Qwen family; sizes from the HF repo list (all created 2025-09-23).
Sources
- qwenlm.github.io/blog/
- huggingface.co/api/models?author=Qwen&sort=createdAt&direction=-1&limit=60&skip=60
- arxiv.org/abs/2510.14276
- qwenlm.github.io/blog/qwen3guard/
This record was checked against its sources on 6 October 2026. How we check