Qwen3Guard-Gen-4B
Alibaba Cloud
A generative safety-moderation model from the Qwen team, fine-tuned from Qwen3-4B, which frames safety classification as an instruction-following task. It labels prompts and responses across three severity tiers - Safe, Controversial and Unsafe - and names the violated categories from a taxonomy of Violent, Non-violent Illegal Acts, Sexual Content or Sexual Acts, Personally Identifiable Information, Suicide & Self-Harm, Unethical Acts, Politically Sensitive Topics, Copyright Violation, and Jailbreak (input only). For responses it additionally emits a Refusal label. The Qwen3Guard series was trained on 1.19 million prompts and responses labelled for safety and supports 119 languages and dialects.
Strengths
- Three-tiered severity classification - Safe, Controversial and Unsafe - instead of a binary verdict, so borderline content can be routed rather than blocked
- Nine named safety categories, with Jailbreak applied to inputs only
- Response moderation also returns a Refusal label, usable for measuring over-refusal
- Supports 119 languages and dialects, per the vendor's model card
- Permissive Apache-2.0 licence with the licence file shipped in the repository
- Part of a matched series: Gen at 0.6B/4B/8B plus a Stream variant with a token-level head for moderating text as it is generated
- Documented serving paths for SGLang and vLLM behind an OpenAI-compatible API
Caveats
- The vendor ships no benchmark numbers as text - the model card's only performance evidence is a chart image (performance.png), so its 'state-of-the-art' claim cannot be checked from the page
- No context window is stated in prose; 32768 appears only inside the SGLang and vLLM launch commands, which is a serving flag rather than an advertised limit
- 'Politically Sensitive Topics' is a vendor-defined category with no equivalent in the MLCommons or OpenAI taxonomies, which may not match a deployment's policy
- The 119-language claim is not broken down per language, and the card's own performance wording singles out only English and Chinese
- The card contains a copy-paste error: the response-moderation example loads 'Qwen/Qwen3Guard-4B-Gen', which is not the repository id
- The Gen variant classifies complete texts; moderating partial output as it streams needs the separate Qwen3Guard-Stream model
- No release date is stated on the model card
- Vendor
- Alibaba Cloud
- License
- Apache-2.0
- Modalities
- text
Capabilities
Resources
Pricing
Open weights under Apache-2.0, self-hosted at no licence cost. The vendor documents serving it with sglang>=0.4.6.post1 or vllm>=0.9.0 behind an OpenAI-compatible endpoint, and with transformers>=4.51.0 locally.