Qwen3Guard: Streaming Safety for Real-Time Tokens

Alibaba Qwen has released Qwen3Guard, the first safety guardrail model in the Qwen family, built on Qwen3 foundation models and fine-tuned for safety classification. It provides precise safety detection for prompts and responses with risk levels and categorized classifications. Qwen3Guard achieves state-of-the-art performance on safety benchmarks across English, Chinese, and multilingual environments. The model comes in two variants: Qwen3Guard-Gen, a generative model for offline safety annotation and filtering, and Qwen3Guard-Stream, which enables efficient real-time streaming safety detection during response generation. Both variants are available in three sizes: 0.6B, 4B, and 8B parameters. Qwen3Guard features real-time token-level streaming detection using lightweight classification heads on the transformer's final layer, a three-tier severity classification (Safe, Unsafe, Controversial) enabling flexible safety policies, and support for 119 languages and dialects. The models are open-sourced on Hugging Face and ModelScope, and are also available via Alibaba Cloud AI Guardrails service.