AI Potluck
Contribute
Back to Gap Map Product / UX / Safety & Guardrails

Qwen3Guard

Alibaba Cloud
open weights / Overall score: 3.0

Qwen3Guard is Alibaba's family of safety moderation models built on Qwen3, in 0.6B, 4B and 8B sizes, that classify prompts and responses as safe, controversial or unsafe across 119 languages. A streaming variant classifies each token as it is generated, for real-time moderation. The models run on vLLM, SGLang and Hugging Face Transformers.

Openness

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

All three Qwen3Guard sizes are released under Apache-2.0 and can be downloaded without gating. The model card describes the safety-labeled training set but does not publish it, which makes this an open-weights release rather than a fully open model.

  • https://huggingface.co/api/models/Qwen/Qwen3Guard-Gen-0.6B recorded 2026-09-29

    The Hugging Face API record for "Qwen/Qwen3Guard-Gen-0.6B" carries the tag "license:apache-2.0", is ungated, and lists a single weight file with no dataset tag; its tags include "base_model:finetune:Qwen/Qwen3-0.6B".

  • https://huggingface.co/api/models/Qwen/Qwen3Guard-Gen-4B recorded 2026-09-29

    The Hugging Face API record for Qwen/Qwen3Guard-Gen-4B carries the tag "license:apache-2.0", is ungated, lists three weight shards ("model-00003-of-00003.safetensors") and carries no dataset tag.

  • https://huggingface.co/api/models/Qwen/Qwen3Guard-Gen-8B recorded 2026-09-29

    The Hugging Face API record for Qwen/Qwen3Guard-Gen-8B carries the tag "license:apache-2.0", is ungated, and lists five weight shards ("model-00005-of-00005.safetensors"); its tags name the base model and paper but no dataset.

  • https://huggingface.co/Qwen/Qwen3Guard-Gen-8B/raw/main/README.md recorded 2026-09-29

    The Gen-8B raw model card front matter declares "license: apache-2.0" and lists no datasets key; the body describes the training corpus only as "trained on a dataset of 1.19 million prompts and responses labeled for safety" without linking or publishing it. Searched the body for fine-tuning and training code: none.

Adoption

2 high confidence
2.0

Adoption is read from Hugging Face downloads for the 8B checkpoint, Qwen/Qwen3Guard-Gen-8B, the only artifact counted here. No other checkpoint's downloads are recorded, so the reading reflects that one size rather than the full family.

Capability

4 medium confidence
4.0

Each verdict also names which of nine harm categories applies, including jailbreak attempts on prompts, and response checks report whether the model refused. The streaming variant reads Qwen3 token IDs, so pairing it with a model on a different tokenizer means re-tokenizing the text first.

  • https://huggingface.co/Qwen/Qwen3Guard-Gen-8B/raw/main/README.md recorded 2026-09-29

    The Gen-8B card says "The series includes models of three sizes (0.6B, 4B, and 8B)", grades verdicts as "safe, controversial, and unsafe severity levels", states "Qwen3Guard-Gen supports 119 languages and dialects", lists nine harm categories ending in "Jailbreak (Only for input)", and shows response checks returning "Refusal: Yes".

  • https://huggingface.co/Qwen/Qwen3Guard-Stream-8B/raw/main/README.md recorded 2026-09-29

    The Stream-8B card says the streaming variant "incorporates a token-level classification head for real-time safety monitoring during incremental text generation" and notes that "Streaming detection requires streaming token IDs as input".

Verified 2026-09-27