Qwen3Guard Technical Report (huggingface.co)

Qwen3Guard is a new suite of multilingual safety guardrail models that tackles two practical limitations of existing safety classifiers: the usual binary safe/unsafe output and the need to see a complete model response before flagging problems. The project offers two specialized variants — Generative Qwen3Guard, which reframes safety assessment as an instruction-following task to produce fine-grained tri-class labels (safe, controversial, unsafe), and Stream Qwen3Guard, which adds a token-level classification head to monitor output as it’s generated. Both types are built to operate with low latency during streaming inference, enabling earlier interventions and reducing exposure to harmful partial outputs. Technically, Qwen3Guard ships in three sizes (0.6B, 4B, 8B parameters) and supports up to 119 languages and dialects, targeting global deployment scenarios. Evaluations show state-of-the-art performance on English, Chinese, and multilingual prompt/response safety benchmarks. The combination of instruction-following triage and token-wise real-time monitoring makes these models easier to tune to domain-specific safety tolerances and compatible with modern streaming LLM pipelines. All models are released under an Apache 2.0 license for public use, offering practitioners a scalable, low-latency option for integrating nuanced safety moderation into production LLM systems.