Qwen3Guard Gen 4B

Qwen3Guard-Gen-4B is a 4B parameter model hosted on Hugging Face that evaluates the last user query in a conversation for content safety. It classifies the query as safe, unsafe, or controversial according to an explicitly defined safety policy and, when unsafe content is present, identifies its category from a fixed list.

The model follows a structured prompt template that instructs it to first determine whether the query belongs to safe or unsafe content based on the policy. Safe queries are those that contain no prohibited content or content that could lead to prohibited content. Unsafe queries contain prohibited content. Controversial queries do not explicitly contain prohibited content but carry intent, context, or potential for misuse that could generate unsafe content under certain conditions. When unsafe content is detected the model also assigns one of the listed categories: Violent, Non-violent Illegal Acts, Sexual Content or Sexual Acts, PII, Suicide & Self-Harm, Unethical Acts, Politically Sensitive Topics, Copyright Violation, or Jailbreak.

It is distributed as an open model on the Hugging Face platform. The provided template shows it is designed to process conversation history ending with a user message and to output a safety assessment. This positions it as a specialized foundation model for safety classification that can be incorporated into larger language-model pipelines.