Meet Qwen3Guard: The Qwen3-based Multilingual Safety Guardrail Models Built for Global, Real-Time AI Safety

Can safety keep up with LLMS in actual time? The QWEN team in Alibaba believes that, and QWEN3Guard has just been charged with a moderate multi-language family designed for moderate demands and the flow of responses at the present time.

QWen3Guard comes in two types: QWEN3Guard-Gen (Toulini workbook reads the context of the claim/full response) and QWEN3Guard-Stream (A workbook is created at the distinctive symbol level that is run as a text). Both were launched in 0.6B, 4B and 8B The sizes of the targeted global parameters and pamphlets with coverage 119 languages ​​and dialects. Open source models, with embracing weights and github repo.

What is new?

• Moderation head flow: Attachment Lightweight rating heads To the final transformer layer – one of them monitors the user’s mentor, the other degrees that were created each symbol in the actual time as Safe / controversial / unsafe. This can enforce policy while producing the response, rather than liquidating after the designated.
• Three -level risk connotations: Beyond safe/inaccurate bilateral signs, a controversial Tier supports adjustable stricter (tightening/diligence) via data and policies – useful when the “border line” content should be directed, not simply dropping it.
• The structured outputs of the general: The obstetric variable belongs to a record head –Safety: ...and Categories: ...and Refusal: ...This is trivial to analyze the pipelines and RL bonus functions. Includes categories Violent non -violent works, sexual content, PII, suicide and self -connection, unethical actions, political sensitive topics, violation of copyright, prison break.

Standards and Safety RL

QWEN research team appears The latest F1 Via English, Chinese and multi-language safety standards for both instant classification and response, with data drawing for QWEN3Guard-Gen against previous open models. While the research team confirms relative gains instead of one compound scale, consistent bullets via settings are the main point.

To train assistants in the direction of the river, the search team that is moved by the safety of the safety uses QWEN3Guard-Gen as a bonus sign. A Only guard The bonus increases safety, but the nails and scratches are slightly Arena-Hard-V2 winning rate; A hybrid The reward (punishing excessive curvature, mixing quality signals) raises the safety degree measured by WildGuard from ~ 60 to> 97 Without thinking tasks, even Hard-HD-V2 preparations up. This is a practical recipe for the teams that witnessed the collapse of the previous reward for “rejecting everything.”

Where it fits?

Most of the open guard models are only classified as complete outputs. QWEN3Guard’s Double heads + registration of the distinctive symbol Compatible with the production agents that flow on responses, and enable Early intervention (The mass, transmitted, or re -guidance) with a decrease in the cost of cumin from re -coding. the controversial Tier also sets a clean setting on the Foundation Policy Holds (for example, “controversial” dealing as unsafe in the organizational contexts, but allowing a review of consumer chat).

summary

QWEN3Guard is a practical training staple: open weights (0.6B/4B/8B), two operating modes (full GEN, the Distinguished Time Stream), signs of the three risks, and multi -language coverage (119 languages). As for the production teams, this is a reliable basis line to replace the filters after allocated in the actual time and for the assistants to the safety bonuses while monitoring the rates of rejection.

verify paperand Jaytap page and Full collection on HF. Do not hesitate to check our GitHub page for lessons, symbols and notebooks. Also, do not hesitate to follow us twitter And do not forget to join 100K+ ML Subreddit And subscribe to Our newsletter.

Asif Razzaq is the CEO of Marktechpost Media Inc .. As a pioneer and vision engineer, ASIF is committed to harnessing the potential of artificial intelligence for social goodness. His last endeavor is to launch the artificial intelligence platform, Marktechpost, which highlights its in -depth coverage of machine learning and deep learning news, which is technically sound and can be easily understood by a wide audience. The platform is proud of more than 2 million monthly views, which shows its popularity among the masses.