AI Potluck
Product / UX / Safety & Guardrails

Qwen3Guard

Alibaba Cloud

Alibaba's Qwen3-based safety moderation model family (0.6B / 4B / 8B) that classifies prompts and responses as safe, controversial, or unsafe across 119 languages, with a token-level streaming variant for real-time moderation. Deployable via vLLM, SGLang, and HF inference.

Qwen3Guard ships in three sizes (0.6B, 4B, 8B) and two variants, Qwen3Guard-Gen for instruction-style classification and Qwen3Guard-Stream for token-level moderation, across 119 languages. The card names a 1.19 million-sample training set but does not publish it. Verified 2026-08-13 via the HF model card.

Openness

3 medium confidence
3.0
weights
open(Qwen3Guard 0.6B/4B/8B on HF)
data
not-released
license
Apache-2.0(OSI)

Open weights under Apache-2.0, an OSI licence, but neither the post-training data nor the fine-tuning recipe is released: the card records the training data as withheld and publishes no recipe. Permissive weights over a closed recipe score 3 here as they do everywhere else on the map, because the rung above requires both the data and the pipeline. Coverage is unusually broad even so - 119 languages, plus a streaming variant.

  • https://huggingface.co/Qwen/Qwen3Guard-Gen-8B recorded 2026-08-13

    Card front matter records license - apache-2.0 with a license_link to the repo's own LICENSE; the repo is ungated and ships five safetensors shards. The body says the series is 'trained on a dataset of 1.19 million prompts and responses labeled for safety' but publishes no such dataset and declares none in cardData.

  • https://huggingface.co/Qwen/Qwen3Guard-Gen-8B/raw/main/README.md recorded 2026-08-13

    Raw card front matter - license - apache-2.0, base_model - Qwen/Qwen3-8B. No datasets key, and the training corpus is described by size only.

Adoption

2 high confidence
2.0

39,581 downloads in the trailing 30 days for Qwen/Qwen3Guard-Gen-8B, which bands at 10K-100K, level 2 on the software and model adoption scale.

Capability

4 medium confidence
4.0

Among the broadest open guardrails on language coverage, with a streaming mode that few others offer.

  • https://huggingface.co/Qwen/Qwen3Guard-Gen-8B recorded 2026-08-13

    Card lists three sizes (0.6B, 4B, 8B) and two variants, Qwen3Guard-Gen and Qwen3Guard-Stream; 'Qwen3Guard-Gen supports 119 languages and dialects'; verdicts are graded safe, controversial and unsafe.

Verified 2026-08-12