Qwen3-8B

Qwen3-8B

Qwen's 8.2B open language model that switches between a thinking mode for hard problems and a plain mode for chat, with tool use and support for over 100 languages.

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: Qwen3-8B is an 8.2B language model from Qwen, part of the Qwen3 generation. One model runs in two modes. A thinking mode works through logic, math and code step by step, and a non-thinking mode answers general questions more directly.

It can call external tools in both modes, and it covers more than 100 languages and dialects. The context length is 32,768 tokens natively and 131,072 with YaRN.

Who it suits: It suits developers and researchers who want a mid-size open model they can run themselves, for reasoning tasks, chat, translation or agent work. A 16.4 GB BF16 file asks for a decent GPU.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a self-hosted model for coding and reasoning | Explaining a stack trace
Customer service | Support teams | Run multilingual chat on their own servers | Replying to a customer in several languages
Research | ML researchers | Study or fine-tune a widely used open model | Comparing thinking and non-thinking answers
Agents | Tool builders | Let a model call external tools | A helper that picks a tool for a request

Skip it if you want a hosted chat service and never run models yourself.

Where to use it:

• Python with the transformers library
• SGLang or vLLM, to serve an OpenAI-compatible endpoint
• Desktop tools such as Ollama, LMStudio and llama.cpp, which the card lists as supporting Qwen3
• MLX-LM and KTransformers, which the card also lists
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Hard reasoning questions | Developers and students | A thinking section followed by an answer
General chat | Product teams | A direct reply in non-thinking mode
Translation and multilingual work | Global teams | Output in many of the 100+ languages

How to set it up: You need a few things in place first.

• Python and pip
• A recent transformers version, since versions below 4.51.0 raise an error about the qwen3 name
• Memory for about 16.4 GB of BF16 weights, or about 4.1 GB at 4-bit by the record's estimate

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Load the model with transformers, or serve it with SGLang or vLLM.

pip install -U huggingface_hub
hf download Qwen/Qwen3-8B

How to use it: In Python, you load the tokenizer and the model, build a chat message, and call generate. The enable_thinking flag in the chat template picks the mode, and it defaults to on. The card then shows how to split the thinking text from the final answer.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-8B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

To serve the model, the card names SGLang version 0.4.6.post1 or newer, or vLLM version 0.8.5 or newer.

python -m sglang.launch_server --model-path Qwen/Qwen3-8B --reasoning-parser qwen3

vllm serve Qwen/Qwen3-8B --enable-reasoning --reasoning-parser deepseek_r1

Example: You ask "Write a short note on what a context window is" with thinking turned on. The model first produces its reasoning, and then a final reply. The card's code separates the two so you can print them apart.

For newcomers and experts: Start with thinking on and one short question, and read both parts of the reply. Then switch the flag off for a quicker, plainer answer.

The model has 36 layers with grouped query attention, using 32 heads for queries and 8 for keys and values. Of the 8.2B parameters, 6.95B are non-embedding. For long inputs, the card points to YaRN, which stretches the context from 32,768 to 131,072 tokens.

Title: Context length of Qwen3-8B
Native context: 32768 tokens
With YaRN: 131072 tokens

Download or load the model
Pick thinking or non-thinking mode
Send a chat message
Read the reasoning and the reply

Good to know:

• License: Apache 2.0
• Cost to run: none named, but you need hardware for about 16.4 GB of BF16 weights
• Data it reaches: it downloads files from Hugging Face onto your computer

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 8.2B
• Weights: about 16.4 GB in BF16; about 4.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 6,269,141
• Likes: 2,184
• License: apache-2.0
• Based on: Qwen/Qwen3-8B-Base
• Quantized versions: 520
• Fine-tunes: 2,330
• Tags: transformers, safetensors, qwen3, text-generation, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3-8B
• Files: https://huggingface.co/Qwen/Qwen3-8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3-8B

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