Qwen releases Qwen3Guard, multilingual moderation models in 0.6B, 4B and 8B
Apache-2.0 safety classifiers covering 119 languages with Safe, Controversial and Unsafe levels; Stream variants are also on the Hub.
For context: Qwen3Guard predates the twelve-month window by a few weeks, and is included because it is a base case for open guard models. The Hugging Face repositories were created on 23 September 2025, and the card cites a technical report published on 16 October 2025.
The 8B card describes a safety moderation model built on Qwen3 that frames classification as an instruction-following task. It moderates user prompts and model responses. The series comes in three sizes, 0.6B, 4B and 8B, and the Hub also lists streaming variants. The licence is Apache-2.0. The card says it was trained on about 1.19 million labelled prompts and responses and supports 119 languages and dialects.
Severity levels are Safe, Controversial and Unsafe. Categories are Violent, Non-violent Illegal Acts, Sexual Content or Sexual Acts, Personally Identifiable Information, Suicide and Self-Harm, Unethical Acts, Politically Sensitive Topics and Copyright Violation, with Jailbreak applying to inputs only.
The card claims state-of-the-art performance on safety benchmarks across English, Chinese and multilingual tasks, but gives no specific scores on the page. The chart is not reproduced here.
The Controversial middle band is the useful part: it lets a team set its own line. The claims are the vendor's.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache-2.0 weights; Hub shows hundreds of thousands of downloads for the 0.6B.
- Useful 4: Three-level severity plus categories for prompts and responses, in a small footprint.
- Notable 3.5: Small multilingual guard family with heavy Hub downloads.
- Novel 3.5: A three-tier severity label rather than binary.
Comments