Qwen3Guard-Stream-0.6B

Qwen3Guard-Stream-0.6B

Checks text for safety token by token while a language model is still writing, rating each part safe, controversial or unsafe in 119 languages and dialects. Suits developers adding live moderation to chat or generation apps.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3Guard-Stream-0.6B is a safety moderation model from the Qwen team, built on Qwen3-0.6B. It carries a token-level classification head, so it can judge text while a reply is still being generated instead of waiting for the finished answer. It rates content as safe, controversial or unsafe.

Who it suits: Developers who run a chat or text generation product and want live safety checks, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | App developers | Watch a model reply as it streams and flag problems early | Stop a reply that turns unsafe midway
Customer support | Support platforms | Screen what a chatbot says to customers | Check each streamed answer
Education | Learning apps | Keep a tutor bot inside safe topics | Flag an unsafe turn in a lesson chat
Global services | Multilingual products | Moderate in 119 languages and dialects | Review replies in several languages

Skip it if you only need to check finished messages, since the series also has Qwen3Guard-Gen, which treats safety as an instruction-following task.

Where to use it: It sits beside a language model that streams its output.

• Python with the Hugging Face transformers library
• Pipelines where the main model shares the Qwen3 tokenizer
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
Prompt check | Developers | A risk level for the whole user message
Streaming reply check | Product teams | A risk level for each new token as it arrives
Category labels | Safety teams | The risk category when a turn is not safe

How to set it up: You need Python and a recent transformers release.

• Python with pip
• transformers version 4.55.0 or newer
• A main model that uses the Qwen3 tokenizer, or a re-tokenizing step for other models
• Memory for the weights: roughly 1.2 GB at BF16 precision, or roughly 299 MB when squeezed to 4-bit

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Update transformers to 4.55.0 or newer.
4. Load the model in Python with trust_remote_code switched on.

pip install -U huggingface_hub
hf download Qwen/Qwen3Guard-Stream-0.6B

How to use it: Load the tokenizer and the model with trust_remote_code set to True, since the Stream architecture needs it. Format the conversation with the chat template, then call stream_moderate_from_ids on the whole user prompt first and on the assistant tokens one at a time. The loading lines look like this.

import torch
from transformers import AutoModel, AutoTokenizer

model_path="Qwen/Qwen3Guard-Stream-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

Tool | What it does
stream_moderate_from_ids | Rates the tokens you pass and returns a result and a stream state
risk_level | Part of the result that says Safe or a higher risk level
category | Part of the result that names the risk category
stream_state | Carries the conversation context from one call to the next

Example: You pass a user message and the start of an assistant reply. The model first rates the user message, then rates each assistant token as it arrives, and your code prints the risk level and, when it is not safe, the category.

For newcomers and experts: If you are new to moderation models, treat this one as a smoke detector for text that is still being written. Run the card's sample once, and watch how the risk level changes as tokens arrive.

If you know the field, the model adds a token-level classification head to a Qwen3 base, so judgments can update during incremental generation. It was trained on 1.19 million labelled prompts and responses. Because streaming works on token IDs, models with another tokenizer need their text re-tokenized into the Qwen3 vocabulary and fed in step by step.

Download files with hf
Load the tokenizer and model
Rate the full user prompt
Rate each assistant token as it streams
Read risk level and category

Good to know: The series comes in 0.6B, 4B and 8B sizes, in a Gen variant and a Stream variant. This repository is the Stream variant at 0.6B.

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the text you send it and the model files on your computer, and nothing else the card names
• Not for: models with a different tokenizer unless you re-tokenize their text into the Qwen3 vocabulary

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: transformers
• Parameters: 597M
• Weights: about 1.2 GB in BF16; about 299 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,581
• Likes: 55
• License: apache-2.0
• Based on: Qwen/Qwen3-0.6B
• Quantized versions: 4
• Fine-tunes: 4
• Tags: transformers, safetensors, qwen3, feature-extraction, text-embeddings-inference

Links

• Model: https://huggingface.co/Qwen/Qwen3Guard-Stream-0.6B
• Files: https://huggingface.co/Qwen/Qwen3Guard-Stream-0.6B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3Guard-Stream-0.6B

Then follow the usage in its model card. Nothing in this kit runs by itself.