Alibaba's Qwen team launches Qwen3Guard, a safety classifier for live AI outputs

Alibaba's Qwen team has released Qwen3Guard, the first dedicated safety guardrail model in its lineup, designed to classify risks on both incoming prompts and generated responses in real time. Built on the Qwen3 foundation models and fine-tuned specifically for safety classification, the system delivers risk levels and categorized classifications aimed at helping developers moderate AI interactions more precisely.

The model claims state-of-the-art performance on major safety benchmarks across English, Chinese, and multilingual environments, according to the Qwen team. For builders shipping LLM-powered products, a guardrail model that scores both sides of the conversation — not just outputs — addresses a real gap: most moderation tools catch bad responses after the fact, but Qwen3Guard promises to flag dangerous prompts upstream as well.

The Qwen3Guard release is the latest in a busy stretch from the team. Over the past two months, Qwen has also shipped Qwen-Image-Edit (an image editing model with text-rendering capabilities), Qwen-Image (a 20B image foundation model), GSPO (a reinforcement learning algorithm designed to prevent model collapse during long training runs), and Qwen-MT (a translation model supporting 92 languages). The team also announced that its blog is migrating from GitHub Pages to a new home at qwen.ai.

Models and weights for Qwen3Guard are available on GitHub, Hugging Face, and ModelScope, with a tech report published alongside the release.