Qwen3-4B-Instruct-2507

Qwen3-4B-Instruct-2507

Answers chat prompts with a 4B model that skips visible thinking and reads up to 256K tokens of context. For developers who want a small open assistant for reasoning, coding and tool use.

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: Qwen3-4B-Instruct-2507 is the 2507 refresh of Qwen3-4B for the setting that skips thinking. The Qwen team made it. It has 4.0B parameters and answers directly, without writing think blocks. The card lists gains in instruction following, reasoning, math, science, coding and tool use.

Who it suits: Developers and students who want a small open model for chat, reasoning and long documents on modest hardware.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a small assistant to an app | Answer user questions inside a product
Education | Students | Practice with math and logic prompts | Work through a reasoning puzzle
Writing | Content teams | Draft and edit text with an open model | Write a short article draft

If you want a model that shows its reasoning steps, you can skip it, since it supports only non-thinking mode.

Where to use it:

• Your own computer with Python and the transformers library
• SGLang or vLLM, which can serve it behind an OpenAI-compatible API
• Qwen Chat, the hosted chat the card links

Use case | Who | What they get back
Chat answers | Anyone | Direct replies without think blocks
Long documents | Analysts | Understanding across up to 262,144 tokens
Tool use | Developers | Function-calling style answers

How to set it up:

• Python with pip
• A recent transformers release, since versions below 4.51.0 fail with a KeyError for qwen3
• About 8 GB of disk space for the BF16 weights, or about 2 GB at 4-bit (an estimate)

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Run the card's Python snippet, which loads the tokenizer and model and prints a reply.

pip install -U huggingface_hub
hf download Qwen/Qwen3-4B-Instruct-2507

How to use it: Send a user message through the chat template, generate, and decode the new tokens. The card's snippet asks for a short introduction to large language models.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-4B-Instruct-2507"

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

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=16384
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

content = tokenizer.decode(output_ids, skip_special_tokens=True)

print("content:", content)

Example: You run the snippet unchanged. The model writes a short introduction to large language models and the script prints it after the label content.

For newcomers and experts: If you are new, run the snippet and change only the prompt. Because this model never writes think blocks, you do not need to set a thinking flag.

For experts, the card lists 36 layers, 32 query heads and 8 key-value heads, with a native context of 262,144 tokens. For serving, it names sglang 0.4.6.post1 or newer and vllm 0.8.5 or newer.

You write a user message
The chat template formats it
The model answers directly, with no think block
You decode the new tokens

Good to know: The card's table compares this model with the older Qwen3-4B non-thinking mode and two others on AIME25.

Title: AIME25 scores from the card
Qwen3-4B-Instruct-2507: 47.4 points
GPT-4.1-nano: 22.7 points
Qwen3-30B-A3B Non-Thinking: 21.6 points
Qwen3-4B Non-Thinking: 19.1 points

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need hardware that can hold the weights.
• Data it reaches: downloading reaches Hugging Face. The snippet does not describe reaching other files or the web.
• Not for: step-by-step reasoning traces. It answers directly and never writes think blocks, so pick another model if you want them.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4B
• Weights: about 8 GB in BF16; about 2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,592,168
• Likes: 1,053
• License: apache-2.0
• Quantized versions: 379
• Fine-tunes: 2,755
• Tags: transformers, safetensors, qwen3, text-generation, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
• Files: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3-4B-Instruct-2507

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