Qwen announced Qwen3Guard as a safety guardrail model line in its model family, built on Qwen3 and fine-tuned for safety classification on both prompts and responses.
The launch adds two deployment modes: Qwen3Guard-Gen for offline or dataset-level annotation and safety-reward workflows, and Qwen3Guard-Stream for token-by-token safety checks while responses are being generated.
The same material says both modes are available as safety-focused model variants, with Qwen3Guard-Stream designed for lower-latency monitoring in live generation pipelines.
Qwen3Guard applies a layered severity policy (safe, controversial, unsafe) so teams can reinterpret controversial outputs by deployment strictness.
The public docs describe the safety-focused dataset and language coverage and release open checkpoints through Hugging Face and ModelScope.
For model operators, this is positioned as a moderation layer for agentic and LLM products: streaming filtering avoids waiting for full-response post-checks, while offline scoring remains available for policy fine control.
